acceptodds
Under review as a conference paper at ICLR 2027

Critical Decision Capacity: Bottleneck vs. Substitutable Degradation under Action Availability

Abstract

Action availability is a structural resource: systems whose useful behaviour depends on conjunctive bottlenecks can collapse abruptly under random loss of that resource, while substitutable structure degrades smoothly. Critical Decision Capacity (CDC) is not merely a name for a failure threshold : it is the CAC-linked critical capacity used as an intervention objective, with estimators , , and optional (coinciding only under stated assumptions). We prove a sufficient-condition dichotomy for path MDPs: sparse digraph bottlenecks admit a directed-bond percolation threshold (sharp-class mechanism), while submodular structure with bounded marginals is equi-Lipschitz (no discontinuous jump). Path-MDP hardening and Leduc blind transfer (frozen rankings under degree-biased / vs-random failure) show the design chain; poker targeted > random fragility remains Tier B, not a competitive-IIG theorem.

Primary area:
reinforcement learning
Keywords:
action availability, critical decision capacity, phase transition, directed bond percolation, structural robustness, reinforcement learning, Markov decision processes, imperfect-information games, submodularity, finite-size scaling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FGYyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This is a difficult paper to read, despite reading it multiple times to understand what it is about. With ChatGPT's help, I think the paper aims to study when an AI agent does not have access to, or is unable to perform certain actions, why some systems would fail completely while others experience relatively gradual performance degradation. The paper suggests the answer depends on whether the system's function depends on a small number of critical decisions (so-called bottleneck systems) or otherwise (substitutable systems). The paper also shows that removing random agent actions is less harmful than targeted action removal (targeted means that some identified important actions are systematically removed). On this end, the paper proposes a concept called Critical Decision Capacity (CDC), which measures how vulnerable the system is to losing available actions to guide which parts of the system should be protected and reinforced.

Critical strengths

None identified.

Critical weaknesses

The paper may contain interesting or novel contributions. However, the lack of clarity, coupled with an overwhelming number of theorems, equations, assumptions, and propositions, makes it difficult to follow the underlying research narrative. I find it hard to appreciate the significance of the proposed contributions. The introduction is too brief to motivate the research. The manuscript reads more like a rebuttal to a previous submission (App. A.13 - Consolidated answers to reviewer questions) + an extension of Kujur 2025 (heavily cited as methodological foundation) than a standalone paper. The repetition and defensive iteration of the arguments from the 1st page (CDC is not merely a name for a failure threshold) and the extensive mathematical justifications without sufficient motivation and intuitive explanation fail to convincingly communicate what the authors aim to establish.

The authors should properly introduce and define all key quantities used throughout the paper. E.g., terms and notations such as (f_c), CAC, and the various CDC measures (structural, value-based, and normalised) are introduced abruptly, even in the abstract, as though they were already familiar to the general readership. Their meanings, interpretations, and distinctions are not adequately explained in the manuscript. Simply citing Kujur 2025 and expecting readers to consult the prior work for essential definitions is not an acceptable substitute for a self-contained presentation. This approach places an unnecessary burden on readers and further undermines the accessibility and clarity of the paper.

Problem setting should also be introduced properly; the paper appears to work on a poker game environment: Kuhn and Leduc poker but the game rules are not properly introduced, Contribution 3 starts right with Leduc-N \in {5,6,7} and mentions a non-existent appear to be development file (algo_multiseed.json). The appendix has many references to development files but no supplementary files included with the submission.

Suspected sign of autoresearch. L132-134: Do not conflate them: CDCstruct and CDCnorm coincide only under Assumption 1 with uniform ρδ on the cut; otherwise report which estimator and why (Prop. 1; Section 6).

Questions for authors

The authors are strongly encouraged to rewrite the manuscript. Jumping directly into mathematical proofs without properly introducing and motivating the underlying research problem is unlikely to convince readers of the significance of the work.

Reviewer_p8ChOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how the value of a fixed policy degrades when a random fraction ff of its actions becomes unavailable. the paper models this as iid bond retention on the action-labelled transition digraph of an MDP or game, with order parameter Φ(f)\Phi(f), the normalized value.

Critical strengths
  • Robustness of agents to loss of available actions understudied compared with observation and reward perturbations.
  • Lemma 1 and Proposition 2 are correct and useful.
  • Proposition 5 gives a simple closed form design law for serial cuts in parralel route
Critical weaknesses
  • The writing is poor, likely written by AI and unedited (most notable is the abstract, the introduction and related works). The reviewer is not against AI writing, but it must be cleaned up so that it is human understandable.
  • In theorem 1: fcf_c is defined by Φ=1/2\Phi=1/2, but theorem 1 locates where Φ→0\Phi\to0. These differ even asymptotically -> fc=1−pcf_c=1-p_c does not follow
  • proposition 3 is wrong. Counter example: a graph with one bond e1=s→We_1 = s\to W and n−1n-1 further bonds that lie on no s→Ws\to W path. For example, a hub hh with edges to dead-end nodes; degree centrality targets these first. Random retention gives Φrand(f)=1−f\Phi_{\rm rand}(f)=1-f, so fcrand=1/2f_c^{\rm rand}=1/2. The centrality rule deletes the n−1n-1 hub edges before e1e_1. Threfore, Φtgt(f)=1\Phi_{\rm tgt}(f)=1 for f<(n−1)/nf<(n-1)/n, giving fctgt=(n−1)/n>1/2f_c^{\rm tgt}=(n-1)/n>1/2 for n≥3n\ge3.
Questions for authors

See weakness.

Reviewer_p232OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1(−1)
Soundness
3
Read review
Paper summary

This paper studies what happens to a decision-making system when some of its available actions are taken away for reasons such as, API rate limits, and regulatory restriction. CAC is a diagnostic which tells you where a system's fragility is concentrated. CDC is meant to answer two follow-up questions: will performance degrade gracefully or hit a cliff as capabilities disappear, and given a limited budget to add redundancy or protect specific decision points, where should that budget go so the protection still works.

The paper proves a sufficient-condition dichotomy for path MDPs: sparse, bottleneck-structured decision graphs admit a percolation-style sharp collapse threshold (Theorem 1), while submodular, substitutable structure with bounded marginals degrades smoothly with no discontinuous jump (Theorem 2). It also gives an exact formula for when two different ways of estimating criticality agree (Proposition 1), and demonstrates the hardening/protection use case on path MDPs and Leduc poker, with the poker results explicitly labeled as empirical/conjectural (Tier B) rather than proven.

The first two paragraphs in related works are easy to follow. But overall, i is hard to understand. For example, CAC is not a widely used academia concept for the meaning in this work. It is helpful to include files like "algo_multiseed.json" for confirmation, but the notation also needs clarification. For example, line 087 problem formulation, it only introduces two agent, then we realized there should be a graph, but the definition of node, edge, actions of each node are all missing.

Critical strengths

The blind-transfer result (Fig. 3) is a meaningful empirical finding: a CDC-based protection ranking, tested under a failure process different from the one used to build it, still beats naive baselines. That portability is a genuinely useful practical signal.

Critical weaknesses

Both Theorem 1 and Theorem 2 are conditional, structural statements about idealized path MDPs and miss the discussion about which regime any actual deployed system falls into. The empirical core (poker fragility, intervention/protection results) is explicitly Tier B and depends on Conjecture 1.

Questions for authors
  1. What in Theorems 1–2 is not already implied by directly citing Newman et al. (2001) for the percolation threshold and standard submodular-optimization smoothness lemmas for the Lipschitz bound?
  2. What is the minimal novel mathematical content beyond the mapping itself?
  3. What principled criterion would a practitioner use to build G for a genuinely new domain?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.