acceptodds
Withdrawn from ICLR 2027

Reliability-Aware Determinantal Point Processes for Robust Informative Data Selection in Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traditional data sampling methods, including those based on Determinantal Point Processes (DPPs), offer deterministic approaches to maximize diversity, assuming that the selected data batches are always available without error. This presumption prohibits their use under probabilistic and disrupted access due to storage outage, imperfect communication, and hardware failures. This problem arises naturally in real-world LLM systems that aggregate data from distributed sources (e.g., RAG pipelines, teacher-student fine-tuning, and LLM-based cooperative perception and planning). We show that the original formulation of DPP-based inference collapses under such conditions. To address this gap, we introduce ProbDPP, a novel reliability-aware implementation of k-DPP that accounts for probabilistic data access by regularizing and decomposing the objective function into a geometric diversity term and unreliability cost, which remains well-posed, facilitating robust batch selection under uncertainty. To cover cases where failure probabilities are not known in advance, we frame this reliability-aware diversity maximization as a combinatorial semi-bandit problem and propose an algorithm that naturally adopts an efficient online learning policy. We evaluate the proposed method on different tasks, including question answering, RAG pipelines, summarization, and image reconstruction, across different benchmarks. Theoretical analysis provides regret bounds for the proposed approach under uncorrelated access links, ensuring performance guarantees.

Primary area:
optimization
Keywords:
Determinantal Point Processes (DPPs), Large Language Model, Data Subset Selection, Robust Optimization under Uncertainty, Combinatorial Semi-Bandits, Online Learning

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_x7BiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProbDPP, a reliability-aware data selection method that combines geometric diversity with source availability under stochastic failures. It derives a regularized objective with an additive reliability term and uses a KL-UCB semi-bandit algorithm to learn unknown availability probabilities.

Critical strengths
  1. The paper addresses a practically relevant selection problem by considering both data diversity and source availability under limited selection budgets.
  2. The evaluation spans QA, RAG, summarization, and visual reconstruction, providing evidence that combining reliability and diversity can benefit multiple downstream applications.
Critical weaknesses
  1. The objective’s connection to received information value is insufficiently justified. Regularization preserves nonzero features for dropped items, retaining full-set geometry while reducing failures to an additive penalty. It therefore neither directly measures surviving information nor distinguishes failure structures with identical marginals.
  2. Each online decision requires a generally intractable subset maximization, yet the regret bound assumes exact optimization. Without a scalable approximation and an analysis of its optimization error, the gap between practical implementation and theoretical guarantees remains unresolved.
  3. The regret proof derives bounds under optimistic confidence estimates but does not explicitly account for confidence-failure events when taking expectations. The required exploration-constant conditions are also unclear, leaving the expected regret guarantee incompletely justified.
  4. The baselines lack joint alternatives such as reliability-aware MMR or expected-surviving-information maximization. Since fixed-cardinality ProbDPP is also equivalent to quality-weighted DPP, the results do not sufficiently isolate the proposed regularization’s added value.
  5. Real failure traces and systematic tests linking availability to content difficulty or source quality are missing. The evaluation also assumes candidate embeddings are available upfront, without accounting for their acquisition costs or how they are obtained when sources are inaccessible.
Questions for authors

See Weakness

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

This paper introduces ProbDPP, an extension of the existing k-DPP method that accounts for data reliability, motivated by the possibility of missing data due to practical constraints. The objective function can be decomposed into a diversity term and an unreliability cost. The paper also formulates a combinatorial semi-bandit problem for reliability-aware diversity maximization. The proposed ProbDPP algorithm demonstrates strong empirical performance across several tasks and provides theoretical guarantees, including an upper bound and a lower bound for a special case.

Critical strengths

This paper identifies a novel and practical problem: diversity-based data selection when the selected data may be unavailable due to stochastic failures, a setting that existing DPP-based methods do not consider. The authors propose a simple and well-posed solution, ProbDPP, which separates the objective into a diversity term and a reliability term, and extend it to the case of unknown reliabilities with a bandit-based online algorithm. The method is supported by theoretical guarantees (logarithmic regret) and by experiments across a wide range of tasks, including question answering, RAG, summarization, and image reconstruction, where it consistently outperforms the baselines.

Critical weaknesses

I'm not an expert on this specific problem, but I want to raise some concerns from my own perspective as a reader. I don't think any of the following would lead to a rejection, but these aspects are just not that strong, in my view.

  1. Presentation. This paper is a bit hard for me to follow. I already know that the determinant can represent the diversity of features, but after reading the first 3 pages of the paper, I still find it hard to understand the significance of this problem. Maybe this is because I'm new to this specific problem, but from my own perspective, there is room for improvement in the writing and presentation. One suggestion is to first introduce the concrete problem with examples, to help readers really understand the importance of the problem itself, and then stress the hardness of the problem. Currently, I cannot get these easily from the writing.

  2. Main contribution. Once the motivation of the problem is clear, the methodology is quite straightforward. The theoretical results by themselves are not enough for an ICLR paper, as the proposed problem is quite simple (one matrix whose log-determinant is −∞) and the algorithm is a very simple modification of a bandit learning algorithm. For the experiments, I remain suspicious, as I don't know whether there are real improvements that help people's understanding. I also find Figure 1 a bit misleading: it normalizes the proposed method to 1, while the other methods do not actually seem to be much worse.

  3. Are the empirical results strong enough? This is my question, since I'm not an expert in this specific domain and all the experiments are new to me. So I remain suspicious about whether the empirical results are really good or just minor improvements. If the related methods (baselines) do not consider some structure of the proposed problem, the proposed method could easily beat all of them while still having little significance, since the proposed method is designed for this newly proposed problem.

Questions for authors

Please address the concerns raised in the Weaknesses section.

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

The paper introduces ProbDPP, a reliability-aware data selection framework that balances diversity and stochastic availability. The authors derive a decomposition of a regularized log-determinant objective and propose a KL-UCB-based semi-bandit algorithm for unknown reliabilities, with logarithmic regret guarantees under stationary assumptions. Experiments on QA, RAG, summarization, and image reconstruction demonstrate improvements over several baselines.

Critical strengths
  1. The decomposition into diversity and reliability terms is mathematically clean, and the online extension provides theoretical regret guarantees under explicit assumptions.

  2. The experimental evaluation covers multiple tasks and consistently demonstrates the benefits of jointly considering diversity and reliability.

Critical weaknesses
  1. Limited methodological novelty. The proposed objective is equivalent to a standard quality-weighted DPP formulation, with reliability entering as an additive modular reward. The paper does not sufficiently establish what is fundamentally new relative to existing weighted DPP methods.

  2. Scalability and practical validation. The regret guarantees assume exact combinatorial optimization, while practical approximate implementations are not theoretically analyzed. Additionally, reliability failures are predominantly simulated, limiting evidence of real-world applicability.

Questions for authors
  1. How does the proposed objective fundamentally differ from standard quality-weighted DPP selection, given the equivalence implied by Equation (29)?

  2. What optimization procedure is used in larger experiments, and how would approximate subset selection affect the regret guarantees?

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

Related papers

Discussion (0)

Sign in to comment.