acceptodds
Withdrawn from ICLR 2027

Rethinking Least-Core Computation in Contextual-Distractor Games

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Game-theoretic attribution explains a model by assigning credit to its features or training examples. The least core has attracted interest as an alternative to Shapley-style averaging because it can expose players that cause substantial harm in rare, high-value contexts. However, least-core allocations are generally nonunique, and the choice of allocation can affect the resulting explanation. In this study, we investigate how payoff selection and coalition sampling affect least-core attribution. Our experiments show that selector choice matters for distinguishing useful and harmful contributions, and that sampling can degrade harmful-player identification across the tested selectors even when useful players remain well identified. These observations motivate efficient computation with all coalition constraints and a well-defined selector. We introduce entropic least core (ELC), a smooth approximation whose unique minimizer follows a continuous path along the temperature to the nucleolus, a classical refinement of the least core. Our experiments show that ELC approximates the nucleolus faster than an LP-based nucleolus solver while retaining small payoff errors, with further GPU acceleration at larger problem sizes. ELC is particularly superior in retrieving and ranking harmful players in the contextual distractors setting.

Primary area:
interpretability and explainable AI
Keywords:
Least Core, Game Theory, Model Explanation

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_4yCFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies the least-core concept, which is a concept from cooperative game theory, that recently is receiving a bit more recognition than before (while still being very much under-explored compared to concepts such as Semivalues like the Shapley or Banzhaf values and their corresponding interactions). This paper studies a phenomenon that appears in games when players are being contextual distractors (i.e. players that usually help little in the payoff, but rarely harm a lot by reducing the payoff a lot). The paper then examines the least-core definition more closely and points out that the selection criterion that is usually used to make the core computable in practice (the core is often an empty set) can be reformulated with a new selection rule. The paper presents this as a new method and shows that this can be approximated with empirical tricks.

Critical strengths
  1. Clearly and transparently written. While I think this paper could be better presented as a whole, it still mostly clearly written and explains the related work and theoretical background well.
  2. The main contribution (equation 9), the ELC is to the best of my knowledge novel and not part of the existing literature. The paper correctly points out that the least-core attribution is not well defined without a selector, and that the usual l2 norm (min norm) selector is an arbitrary default. This is very good. Moreover, Proposition 4.2 shows that the entropy least core is connected to the nucleolus computation algorithm and one can now compute the new ELC variant with a GPU-friendly algorithm via projected gradient descent. This is a worthwhile contribution in an important research area. Particularly, the core is underdeveloped in the XAI literature and making it, or concepts derived from this definition seem very promising for making attributions, or generally explanations, more robust. Albeit, this paper does not make a very convincing case of this, yet.
Critical weaknesses
  1. Limited Novelty. This work here is heavily based on a published workshop paper by Nagarajan et al. (2026). The conceptual basis (definition 3.3, least core vs. Shapley) is already discussed in the workshop paper. The main theoretical results, Proposition 4.2, seems to be a direct application of Alvarez (2000), which is heavily cited in the proof (which of course is good). This would, ultimately, not be a deal breaker if the proposed method would have been better presented and evaluated as a _new XAI method. Which brings me to my main concern.

  2. The main method is underdeveloped. First, the contextual distractor setting (definition 3.3) is a very niche issue. The paper does not bring one real-life scenario of this phenomenon. While I am sure synthetically this may occur and it is interesting to investigate this, however, as it stands now, it is quite hard for an XAI researcher to place this contribution inside the existing literature on Shapley attribution. This is mainly a fact that the evaluation does not really help me understand what the benefit of the new formulation is. While I am sure some exist, I am unsure what implications this method would have for the attribution researcher and what better explanations I can create. Therein, the work feels quite technical and hand-crafted and not really impactful for the greater research community.

  3. The used baselines in the existing experimental setup are quite weak. The Shaply estimator is a very bad estimator with very high variance. While it is kind-of like comparing apples with oranges, the baseline still should be the best it could be. Much more sophisticated approximators exist, such as RegressionMSR [1], LevaragSHAP [2], or OddSHAP [3]. The only downstream, implication of the work seems to be the results of Table 5 (the data valuation experiment), which shows only very small improvements over the baselines.

  4. Finally, as it stands this work also contains a collection of minor, more easier fixable, issues (see below) which taken together also weigh as a larger weakness.

Minor, potentially Fixable Weaknesses:

  • While later sections (e.g. lines 261-267) are very detailed, earlier parts of Section 3, particularly Section 3.1 introducing the general game theory concepts, does not contain any references. The definitions are just stated as such and it is unclear where each representation is coming from. This is mostly an issue for the ϵ\epsilon-core as it is the most "non-standard definition" here.
  • Inconsistent / Imprecise Language: Sometimes the language is not so clear: "min-norm selector" but in the paragraph "minimum norm selector". This inconsistency makes the comprehension a bit harder. "Let D ⊂N be a nonempty group." (line 180). Why use "group" as a stand in for coalition here. Naming the same thing differently is hard to follow.
  • Definition 3.3, the Contextual distractor, is quite central for this paper, yet it receives little to no intuition for the reader. The definition is quite technical and the harmfulness implication in lines 186-187 is also very abstract. This should have been better developed and requires.

Reference

Questions for authors

Suggestions: Here is a small list of minor suggestions for small things that irked me.

  • I think your abbreviation "ELC" for entropic least core is not a good choice. ELC is also known as egalitarian least core (the one with the l2 objective as the selection rule) and I would say this is what people would know ELC to stand for in the game theory literature. If I were you, I'd change the abbreviation.
  • I highly suggest to use \phi as the attribution vector as is, in my opinion, quite standard in the literature. Writing xx feels very weird.
Reviewer_UrvROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tries to improve on the least core attribution method, which is an alternative to shapely for in-game attribution problem. It points out that least core has a few problems. So when the payoff is not unique, the selector used to apply on tie breaks affects explanation, where the commonly used minimum norm flattens differences among useful players. Nucleolus solves these issue but is computationally expensive. The proposed entropic least core replaces maximum access complaint with a temperature controlled log some exp objective, and solved by gradient descent. Paper verifies it on random and UCI games where the ELC is able to approximate the nucleolus significantly faster. They also run evaluation on synthetic contextual destructor games where ELC identifies the destructor player contribution and also orders the positive players better. The more realistic downstream tasks come up with small or mixed gains.

Critical strengths
  1. The question posed by the paper is very relevant. I like the toy example which shows this problem clearly. It matters for my assessment because although the example was synthetic, it highlights an issue which might be happening in real-world datasets at an unknown prevalence. So it seems very important to have a fix for it. The solution is simple and elegant and comes with convergence results. The speedups are also confirmed empirically. So the method leaves reasonably little doubt to successfully motivate its use and trial in small enough production scenarios.
  2. The algorithm seems well tested and the experimental design is very scientific. The authors add several controls for all their evaluations and run parameter sweeps which increases the confidence in the experiments run. The paper also includes several unfavorable results which looks like very transparent reporting. One thing I like about the algorithm is the reliability where the algorithm is bound to return something sensible. It is unlikely to fail numerically or converge in unexpected ways.
Critical weaknesses
  1. Somewhat related to the critical strength 1), the practical scope is there, but is indeed limited to sparse data regimes. The algorithm still enumerates all the coalitions in these settings, so solving the full linear problem is also comparably fast. So, ELC speed advantage is not as profound .
  2. The advantage of this algorithm successfully appears in synthetic games constructed by the paper. The real tasks (which I appreciate the paper including) seem to show small games, which leads me to conclude this algorithm has blind spots of its own which the paper has not sufficiently explored.
Questions for authors

The nucleolus baseline solves one LP per coalition at each step, which is slow. How does ELC compare against a faster nucleolus solver, with both methods run to the same accuracy? Are there another class of games where this is useful? i.e. do the ranking advantages hold for games that aren’t built from additive useful players plus a single bonus?

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

This paper investigates how payoff selection and coalition sampling affect least-core attribution. It proposes ELC which uses log-sum-exp smoothing to construct a surrogate objective and accelerated projected gradient descent to solve it, yielding an approximation to the nucleolus. The paper analyzes the uniqueness of the ELC solution and its convergence to the nucleolus as the temperature approaches zero, and empirically evaluates computational efficiency and attribution performance.

Critical strengths
  1. The paper is clearly organized, and the presentation follows a coherent progression.
  2. ELC is precisely formulated, with theoretical analysis of its uniqueness and convergence to the nucleolus in the zero-temperature limit.
Critical weaknesses
  1. The practical significance of the problem remains insufficiently established. The paper highlights the role of payoff selection in contextual-distractor identification, but the experiments do not clearly demonstrate which practical tasks would benefit substantially from this capability. The authors openly acknowledge this limitation. Nevertheless, because contextual-distractor identification is a central motivation for the paper, the limited evidence of practical benefits still makes it difficult to assess the significance of the contribution.

  2. The attribution benefits of choosing the nucleolus as ELC’s approximation target are not sufficiently validated. The authors motivate ELC by arguing that the nucleolus is the most successful selector for contextual-distractor identification but is computationally expensive. However, in Table 1, the nucleolus and min-norm primarily differ in their ranking of useful players, while assigning identical payoffs to the distractors that are the paper’s central focus. Table 4 does not include an exact nucleolus baseline. Thus, the evidence supports ELC’s ability to approximate the nucleolus, but does not sufficiently establish the attribution advantages of its target selector.

  3. The trade-off between ranking improvements and computational cost remains unclear. In Table 4, ELC and min-norm achieve the same identification accuracy with full coalitions, and ELC’s advantages primarily concern within-group rankings. However, the table does not report the corresponding computational costs. Table 7 compares solver times at larger scales but uses distractors with identical harm levels, preventing an assessment of the advantage in harm ranking. Thus, the experiments do not establish the computational cost of obtaining better harm rankings than min-norm or whether these ranking improvements persist as the problem size increases.

  4. Downstream gains are modest and vary across tasks. In Table 5, ELC’s improvements over min-norm on Gaussian are statistically supported but modest in magnitude. On Dog–fish, the differences are all smaller than 0.1 percentage points in magnitude, and some confidence intervals include zero. The practical significance and consistency of these gains across tasks require further validation.


Formatting Problem. Table 2 is set as a text-wrapped table flush against the right margin, whereas the ICLR formatting instructions require that all tables be centered.

Questions for authors
  1. Under the same game settings, what computational costs accompany ELC’s ranking improvements over min-norm? How does this relationship change with problem size?
  2. Do the improvements in Table 5 primarily arise from more accurate valuation of useful samples or better identification of harmful samples? Can the current results distinguish between these explanations?

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

Related papers

Discussion (0)

Sign in to comment.