acceptodds
Under review as a conference paper at ICLR 2027

FULCRUM: Pair-Dependent Preferences Beyond Transitive Rankings

Abstract

Assigning each option one context-independent scalar score, as reward models and fixed priority orderings do, makes every preference transitive. Judgements on several criteria need not be: when the criterion that decides depends on the pair compared, the relation can cycle and no single ordering represents it. We study FULCRUM, a calculus for structured argumentation in which the priority over comparison criteria is chosen from the pair itself, with the resulting relation converted into defeats under standard semantics. Its endogenous MAX-GAP rule induces strict cycles of every length that no preorder realises, and we characterise exactly, through pairwise disagreement and a minimum weighted feedback arc set, the best accuracy any linear order can attain on a given set of comparisons. For a reasonable ordering in the sense of Modgil and Prakken, closure and consistency hold without transitivity, and the aggregation used to score conjunctions decides whether the relevant reasonableness clause holds. A soft-gated comparator makes the structure learnable with criterion-level attribution. Empirically, in a controlled world whose true rule is gap-driven it gains up to +7.3 accuracy points over a linear Bradley–Terry model, above the transitive ceiling; it is neutral on transitive preferences, and more flexible pairwise comparators perform better when the dependence has a different form.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
preference learning, non-transitive preferences, pairwise comparison, multi-criteria decision making, structured argumentation, neuro-symbolic learning, representational capacity

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

The paper looks at the problem of dealing with non-transitive preferences that arise in multi-criteria comparison methods. It defines a preference-based argumentation approach, where the preference relation is based on pair-wise comparison: the criterion on which two arguments differ most decides. It's demonstrated that this may lead to defeat cycles (e.g. a directed 3-cycle with no non-empty extension) that no preorder-based framework can induce. All this is then modelled in the ASPIC+ formalism, which is shown to work even with non-transitive preferences.

Critical strengths

The paper deals with a potentially interesting problem (intransitive preferences) and takes a formal rigorous approach supported by rationality postulates.

Critical weaknesses

The main weakness of this paper is its presentation. I find the writing style very unpleasant. The impression I get is that the authors relied on an LLM to compress an originally much longer text into something that fits the page limit. It appears that compactness was favored over clarity. Overall, the paper is dense and hard to understand. Many of the results (theorems/propositions/corollaries) are provided without proper motivation or explanation. Theorems and propositions are very wordy, and nearly impossible to check for correctness. I furthermore have serious doubts about whether the format (i.e. a conference paper with the given page limit) is the right form for this work. I also have doubts about the relevance of the paper with respect to the theme of the conference and I think it would fit much better at a KR/argumentation venue.

Questions for authors

none

3 comments
AuthorsOpenReview ↗

Response to kdmw

Thank you for highlighting the difficulty of following the presentation. We have reduced the density of the main text and made the learning question and contribution boundaries more explicit.

Presentation changes

  • The introduction begins with the limitation of one fixed score per option and uses a three-option example to explain pair-dependent comparisons.
  • Section 2.2 introduces the closest prior model before describing our extensions.
  • Formal results in Sections 3 and 4 are introduced by the question they answer and followed by their implications. Three secondary statements and the framework survey have moved to the appendix.
  • Table 1 distinguishes the discrete rule, the structured construction with a conditional guarantee, and the unrestricted fitted comparator.
  • Section 3.1 explains empty acceptance in ordinary terms, and Appendix B restates the source definitions needed for the argumentation results.
  • The limitations separately address theoretical assumptions, dataset scope, and empirical findings.
  • Related work now distinguishes rule priorities, attack reversal, numerical strength propagation, and acceptability rankings rather than treating them as one fixed-ordering mechanism.

Relevance to learning and representation

The paper asks what fixed-utility preference models cannot represent, how much accuracy that restriction can cost on a given comparison set, and when a more flexible learned comparator helps. The revised experiments compare against the closest salient-feature model and generic pairwise alternatives on five datasets. The argumentation component then studies conditions under which cyclic comparisons can be used without violating closure and consistency.

The revised paper does not claim that the unrestricted learned comparator inherits those guarantees, or that its empirical gains are uniform. It also replaces the disputed Theorem 6(a), states the assumptions explicitly, and provides the revised proof and source definitions for inspection.

Dear Reviewer kdmw,

Thanks for your time and effort in reviewing our work. Your comments are valuable to improve the quality of it. As the author-reviewer discussion period will end in about 2 days, we gently request that you read the rebuttal and check whether your concerns are addressed. Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

AuthorsOpenReview ↗

Anymore concern?

As we approach the end of the rebuttal period, we would appreciate knowing whether our responses and the updates to the manuscript have addressed your concerns. If so, we would be grateful if you could reconsider your rating in light of these changes. We remain available to address any remaining questions or concerns. Thank you in advance for your consideration.

Reviewer_8w1KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes FULCRUM, a method for comparing arguments based on the criterion on which they differ most. Since different criteria can matter in different comparisons, the method can produce preference cycles. The authors study which preferences the method can represent and how much accuracy is lost when using a single ranking. They also connect the method to structured argumentation, where arguments are built from premises and rules, can attack one another, and are evaluated to determine which arguments are accepted. They study when their method can be used in this setting without losing consistency guarantees. Finally, they train a model to learn these comparisons and test it on synthetic and real datasets.

Critical strengths

A potential strength of the paper is its connection to structured argumentation. However, I am not very familiar with the argumentation literature, so I find it difficult to assess the novelty and significance of this contribution, particularly to what extent it follows from existing results.

Critical weaknesses

Initially I found the max-gap rule to be a somewhat arbitrary modelling choice, but then I found a paper (Bower and Balzano (ICML 2020), Preference Modeling with Context-Dependent Salient Features.) that already proposes choosing which features matter based on the pair being compared, specifically selecting the features with the largest differences. This seemingly very cosely related work was omitted fromt the paper.

Experimentally, the model performs well on synthetic data designed around its assumptions, but shows limited improvements on real datasets.

The proof of Theorem 6(a) appears to rely on a stronger notion of strict continuation than Definition 17 of Modgil and Prakken (2013), which the paper explicitly cites

Questions for authors

Which of the results are novel? Could the authors clarify which definition of strict continuation they use and provide a proof that their scoring construction satisfies the required assumptions?

3 comments
AuthorsOpenReview ↗

Response to 8w1K

Thank you for identifying the directly relevant work by Bower and Balzano (2020). We have added it to the main discussion, revised the contribution claims, and included matched comparisons.

Relationship to the salient-feature model

Their general model allows a known selection function. Its top-t variant ranks coordinates by two-item variance, which is proportional to the squared difference between the items. Under the same preprocessing and tie-breaking, their top-1 selection therefore agrees with MAX-GAP, and its score is the diagonal-W special case of our hard gate. Their model already represents intransitive preferences. Sections 1 and 2.2 and Appendix G now acknowledge this explicitly; we do not claim pair-dependent selection, the largest-gap rule, or the existence of cycles as new.

The revised contribution claims distinguish four components:

  1. Capacity characterizations for the largest-gap rule (Theorem 4).
  2. Use of the classical optimal-ordering value as a dataset-specific accuracy ceiling. Theorem 5 is presented as an accuracy restatement of the weighted optimal-ordering objective associated with Kemeny aggregation, with Slater's objective as the unweighted tournament case, not a new optimization result.
  3. A material-based scoring construction that satisfies the structured-argumentation conditions, and an analysis of the role of aggregation (Theorems 6 and 7).
  4. Cross-criterion weights and soft selection as extensions of the salient-feature model, evaluated separately through ablations.

Direct experimental comparison

We implemented the salient-feature model on the same splits and standardized inputs, without a position intercept, and selected t on validation data. A second version jointly selects an l2 penalty. We also evaluate native per-item features for Pokémon and Sushi-B. Appendix H.1 documents this matched adaptation rather than claiming exact replication. It uses our optimizer, regularization grid, and mean-absolute scaling; the source uses training-mean/standard-deviation normalization and documents validation-based tuning in its supplement.

Across ten seeds, FULCRUM-soft exceeds validation-selected top-t by 6.0 ± 0.7 accuracy points in World C, 3.8 ± 1.6 in World M, 1.6 ± 0.1 on Pokémon, and 0.9 ± 0.2 on HelpSteer2, with no improvement on Sushi-B (Tables 2 and 17). The reported intervals summarize paired per-seed differences. The ablations associate the World C gain with cross-criterion weights and the World M gain with soft selection.

A simpler learned-scale top-1 variant performs better than FULCRUM-soft in World M by 3.2 ± 2.1 points and on Pokémon by 1.7 ± 0.1 points. We report these results and narrow the claim accordingly: the full comparator is not uniformly better than simpler salient-feature variants.

Limited gains on empirical data

We agree that the empirical gains are limited. On Sushi-B, FULCRUM-soft's paired interval against Bradley-Terry includes zero. HelpSteer2 has only 352 test pairs, so the gains correspond to a few predictions; without the position intercept, FULCRUM-soft and BT-MLP reach 0.920 and 0.918. Generic pairwise models outperform FULCRUM-soft on Pokémon. The abstract, Section 5.2, and the limitations now reflect these results.

Theorem 6(a)

The original proof used a stronger notion of strict continuation than Definition 17 permits. Unary strict rules do not exclude the reviewer's continuation, and last-rule scoring is not invariant under singleton continuation. We have replaced this construction with scoring over all fallible material. Section 4 states the four source conditions, and Appendix A.7 proves the revised reasonableness result, including overlapping material and mixed fallibility cases. The general result uses one shared inclusion-monotone map; the cycle witness is explicitly qualified by maximum or nonnegative-sum aggregation. This is a substantive correction, not a reinterpretation of the original theorem.

Dear Reviewer 8w1K,

Thanks for your time and effort in reviewing our work. Your comments are valuable to improve the quality of it. As the author-reviewer discussion period will end in about 2 days, we gently request that you read the rebuttal and check whether your concerns are addressed. Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

AuthorsOpenReview ↗

Anymore concern?

As we approach the end of the rebuttal period, we would appreciate knowing whether our responses and the updates to the manuscript have addressed your concerns. If so, we would be grateful if you could reconsider your rating in light of these changes. We remain available to address any remaining questions or concerns. Thank you in advance for your consideration.

Reviewer_zng9OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a rule for comparing arguments scored on several criteria. For each pair, the criterion with the largest difference determines the preference, unless the CERTAIN/PROBABLE distinction takes precedence. Since the deciding criterion can vary from pair to pair, the resulting preferences can form cycles. The authors use these preferences directly to turn attacks into defeats in ASPIC+, then apply the standard argumentation semantics to the resulting graph.

The paper makes three main contributions. Representation. The example in Figure 1 shows that no single ordering can reproduce all the pairwise preferences in the three-argument example. Theorem 5 also gives an exact upper bound on the accuracy any single ordering can achieve on a given set of comparisons. Rationality. The reasonableness conditions of Modgil and Prakken do not require transitivity. The corresponding closure and consistency postulates can therefore still hold with cyclic preferences. Theorem 7 further shows that this depends on how conjunctions of arguments are scored. Learning. Since the theoretical results concern the discrete rule, the authors introduce a learnable relaxation with a soft gate. Its sharpness is fitted to the data, allowing it to range from a linear utility model to the original hard rule. The experiments evaluate this learned comparator rather than the discrete rule itself. The evaluation covers two synthetic settings and three datasets (Pokémon, Sushi-B and HelpSteer2), with comparisons against both transitive and cycle-capable baselines. The exact upper bound from Theorem 5 is also reported alongside the model results. The learned comparator exceeds this bound only in the synthetic setting where the labels are generated by the proposed rule, as the authors acknowledge.

Critical strengths

S1: I found Section 4.2 particularly interesting, as it establishes a useful result about ASPIC+ postulates. The authors build on the corrected results of Modgil and Prakken (2013) and specify the conditions under which a FULCRUM framework satisfies the relevant postulates. A key observation is that the required reasonableness condition does not impose transitivity on the preference ordering. This means that closure and consistency can still hold even when preferences are cyclic. I had not seen this possibility explored in this way before. Theorem 7 then identifies a further factor that determines whether the postulates are preserved: how conjunctions of arguments are scored. The striking point, to me, is that the usual ASPIC+ choice, weakest link, is precisely the one that fails. I think this is a useful result for anyone instantiating ASPIC+ with multi-criteria scores. It shows that the choice of aggregation matters in its own right, alongside the rule used to compare arguments. It also provides a clear contribution that stands independently of some of my concerns about the rest of the paper.

S2: I also appreciated the theoretical upper bound established in Theorem 5 and the care taken in reporting the experiments. By plotting the exact ceiling alongside the fitted models (Figure 2(a)), the authors make it possible to distinguish errors that are unavoidable for any single ordering from those due to the model itself. The protocol is thorough, and the ablations of Section 6.1 in particular tie the gain to the gap-based selection rather than to model capacity. The authors also report cases where their inductive bias offers no advantage, rather than suggesting that it works equally well across settings. The experiments are convincing in their design, even where the results themselves are mixed.

Critical weaknesses

W1: The proofs seem for the most part rigorously written. Theorem 6(a), however, puzzled me, because I would tend to think that an example such as the following is allowed by the definitions: with an axiom L, defeasible rules L => A, L => B, L => C and strict rules A -> L, B -> L, C -> L, the chain L => B -> L => C looks to me like a strict continuation of {b, c} in the sense of Definition 17, although every strict rule has one antecedent, and under the score rule of Section 4.2 the set clause would then fail on the witness. I may well have missed something, but if not, would this not mean that the statement needs an additional assumption, for instance that no formula is derived twice within an argument? See Q1. More generally, I would advise the authors to recall, at least in an appendix, every definition they rely on directly: Definitions 5, 16 and 17 of Modgil and Prakken are used but not restated, or only in part, and their absence leaves the reading, and therefore the checking of Section 4.2, incomplete. This matters because the rationality part is what the paper presents as its original contribution.

W2: My second concern is the gap between the theoretical results and the experimental evaluation. Since the learned comparator, as the authors report, outperforms generic pairwise models only where the labels follow its rule, the case for it rests on what it adds beyond accuracy, namely criterion-level attribution, which Section 6.3 tests, and the argumentation layer, by which Section 2 distinguishes FULCRUM from pairwise preference models ("downstream use (defeat semantics with postulate guarantees)"). Yet that layer is the part the experiments leave aside. The introduction makes clear that the learned comparator is a relaxation of the discrete rule, and the experiments focus on the comparator rather than the rule itself. The latter appears only as a fixed MAX-GAP baseline, which performs worse than BT even on World M and whose results on the empirical datasets are not discussed. More importantly, the argumentation layer itself does not seem to be tested on data. Beyond the example in Section 4.1 and the checks accompanying the theorems, no extensions appear to be computed on data-derived conflict sets, either with the discrete rule or with the fitted comparator. See Q2 and Q3.

Questions for authors

Q1: Is Theorem 6(a) meant for arguments in which no formula is derived twice, as the forward-chaining construction mentioned in Appendix B suggests? If so, would the statement be amended accordingly? Relatedly, should the postulate conditions be read as concerning the discrete rule only, or do some of them hold for the fitted comparator with a general W?

Q2: Have extensions been computed on any conflict set beyond the witness, under MAX-GAP or under the fitted comparator, and if so, what do they show that a single ordering would not? Relatedly, how should the fixed MAX-GAP rule's behaviour on the empirical datasets be read with respect to the discrete rule the theory studies?

Q3: On the witness, Section 4.1 reports an empty grounded extension, the empty set as the only preferred extension and no stable extension under the rule, whereas the indifferent preorder has three preferred extensions. What should a reasoner conclude? Is the empty preferred extension the intended behaviour when each argument dominates another on a different criterion, or is the separation of Theorem 2 meant only as a difference between graphs?

4 comments
AuthorsOpenReview ↗

Response to zng9 - Q1

Thank you for identifying the strict-continuation issue and for the three questions. We address them separately below.

W1 and Q1: Theorem 6(a) and the definitions

Your counterexample is valid. Under Definition 17, the chain L ⇒ B → L ⇒ C is a strict continuation of {b, c}: it uses the required ordinary premises and defeasible rules, and its strict-rule and axiom sets contain the required unions. Unary strict rules therefore do not justify the inference in our original proof.

The original last-rule score also fails the singleton-continuation condition. For example, L ⇒ B → L ⇒ C and its continuation L ⇒ B → L ⇒ C → L ⇒ B use the same fallible material but receive different last-rule scores. We have replaced the construction rather than adding an assumption that was absent from the submission.

In the revised construction, every ordinary premise and defeasible rule has a score vector. A fallible argument receives the coordinatewise maximum over all the fallible material it uses. Strict continuation preserves this material for a single argument and takes its union for a set of arguments. With MAX-GAP, a fixed criterion tie-break, and priority for strict-and-firm arguments, Theorem 6 establishes the reasonableness conditions for strict rules of any arity. Appendix A.7 covers overlapping material and mixed sets of strict-and-firm and fallible arguments. The general proof uses one shared inclusion-monotone scoring function. The cycle witness additionally relies on preserving its singleton rule vectors, which maximum aggregation and nonnegative summation do. We now state this qualification explicitly: monotonicity alone could give constant scores and no cycle. Theorem 7 analyzes the set clause and gives counterexamples for coordinatewise-minimum scoring followed by MAX-GAP; these are not counterexamples to standard ASPIC+ weakest-link lifting.

This is a sufficient construction, not a general account of argument strength: adding fallible material cannot lower a criterion score, unlike weakest-link scoring. Combining weakest-link strength with a separate preponderance aggregation remains open. We have not restricted the argument set to derivations without repeated formulas, because the cited postulate results would require a separate justification under that restriction.

Section 4 now states all four conditions of Definition 17 in equation (4). Appendix B restates the source definitions and distinguishes fallible material, maximal fallible subarguments, and the last defeasible rule. The finite checks in Appendix A.7 support the implementation checks but do not replace the proof.

Q1: Does the fitted comparator inherit the guarantee?

No. The guarantee requires the stated material-based scoring construction and discrete comparison rule. An unrestricted learned matrix W need not satisfy it. With W = −I and maximum aggregation, each argument in the three-vector example beats the aggregate of the other two, violating the set clause. A finite learned strength term is also not equivalent to unconditional priority for strict-and-firm arguments.

Table 1 and Section 4 now distinguish these cases. A learned scale and sign applied to each material element before maximum aggregation is one variant covered by the construction; unrestricted hard and soft comparators are evaluated empirically without this guarantee.

AuthorsOpenReview ↗

Response to zng9 - Q2 & Q3

W2 and Q2: Downstream extensions

We added two studies. Neither is presented as a real-world argumentation benchmark, because the preference datasets do not provide arguments and attack relations.

Sampled conflict graphs. We computed exact grounded, preferred, and stable extensions on 10,000 graphs over sets of 6 to 12 held-out items. The study includes complete mutual-conflict graphs and graphs with each conflict sampled with probability one half. Graphs are fixed across comparators and are not selected according to model outputs. A subsample was cross-checked with two independent solvers (Appendix H.3).

Table 3 reports the complete-conflict setting. The generating rule yields an empty grounded extension and no stable extension in 36.5 ± 6.5% of World C sets and 14.4 ± 6.7% of World M sets. A single ordering instead selects a winner in every complete-conflict set. Agreement with the reference grounded extension is 47.8% for Bradley-Terry, 63.5% for FULCRUM-soft, and 90.9% for the large pair-MLP in World C; the corresponding World M values are 68.8%, 86.5%, and 75.5%.

These results show how pairwise errors affect accepted sets. The method ranking follows pairwise accuracy; the study does not establish an additional downstream advantage for the gated architecture.

Structured theories. We also enumerated arguments and complete extensions in 1,600 small synthetic theories with combining strict rules closed under transposition. The construction of Theorem 6 produced no violation of Definition 18 or the tested postulates. Minimum-based scoring produced an inconsistent extension in 76 to 119 of 400 theories per score source; reversing the comparison produced one in 52 to 133 (Table 19). These are finite tests of the construction, not evidence of real-world usefulness.

Q2: Why does the fixed rule perform poorly?

The previously omitted World M row is now included: fixed MAX-GAP reaches 0.785 ± 0.067, below Bradley-Terry (Tables 2 and 8). The diagnostic identifies a criterion-scale mismatch. The generating rule uses hidden scales when selecting the largest gap, whereas the fixed baseline uses standardized observed differences. It selects the generator's criterion on about 64% of test pairs and predicts correctly on about 48% of pairs where the selections differ. Appendix H.2 also identifies polarity-estimation errors in three seeds.

Using the generator's scales reproduces every test label, but this is a privileged diagnostic control, not a deployable baseline. Learning criterion scales from the noisy training data reaches 0.984 ± 0.022 (Table 18). On the empirical datasets, the fixed rule remains substantially below Bradley-Terry: 0.676, 0.616, and 0.815 versus 0.887, 0.656, and 0.921. Sections 5.1 and 5.2 now discuss these limitations directly.

Q3: What should a reasoner conclude?

In the three-cycle, no argument is defended by a conflict-free set. The grounded extension is empty, the empty set is the only preferred extension, and no stable extension exists. Under these semantics, the reasoner commits to none of the three arguments. This represents suspended judgment, not a conclusion that the claims are false. We do not interpret vacuous membership in every stable extension as substantive acceptance.

The indifferent preorder instead permits three alternative singleton preferred and stable extensions. Theorem 2 establishes a difference in acceptance outcomes, not that suspended judgment is preferable in every application. Whether to abstain, present alternatives, or apply a further decision rule depends on the task. Section 3.1 now explains this distinction and clarifies that both preferred and stable semantics support credulous and skeptical queries.

Thank you for the detailed response. My concern about Theorem 6(a) is addressed. Scoring an argument over all its fallible material seems reasonable to me, although the new proof should be carefully examined as part of a complete review of the manuscript.

My second concern is partly addressed. The argumentation layer is now tested in two new studies, and the behaviour of the fixed rule is discussed. However, both studies are synthetic and, as the authors acknowledge, they do not show any practical benefit from using the fitted comparator, which does not inherit the guarantees.

The paper now builds on a substantially revised central result and new findings that still deserve a proper discussion. I believe these changes deserve a full round of review. I therefore maintain my rating for this submission

Thank you for clarifying your assessment and for confirming that the revised construction addresses your concern about Theorem 6(a). We agree that the proof should be examined together with the complete manuscript.

We also agree with your assessment of the empirical scope. The two controlled studies examine how pairwise decisions affect accepted argument sets and how the stated scoring conditions relate to consistency. They do not establish a practical advantage for the fitted comparator, and that comparator does not inherit the theorem’s guarantees. We therefore distinguish the theoretical construction from the learned model and accordingly restrict our conclusions.

We recognize that the changes to the central result and the additional findings warrant careful evaluation. Thank you for identifying both the resolved issue and the remaining gap in evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.