acceptodds
Under review as a conference paper at ICLR 2027

Knowledge transfer between simple discrete worlds

Abstract

Adapting to a new world and reasoning correctly about which outcomes are possible under rules one has not met before is a basic capacity underlying general intelligence. We study it in small discrete worlds: sets of grid configurations consistent with a shared set of rules, concretely Sudoku-like puzzles built from a common base and a handful of extra rules. We train a single depth-recurrent model on many combinations of rules and test it at four levels of novelty: a new puzzle, a new combination of known rules, a pairing of rules never seen together, and a rule instance not seen at all during training. The model receives the active rules either as a symbolic description or as two solved examples, and a model that gets no rule information is used as a baseline. At recombining rules it was trained on, the description matches or exceeds examples ( against on new combinations). On rule instances whose components never co-occurred in training, the description falls below the no-information baseline while examples stay above it. The framework allows us to test many such generalization regimes in more systematic way than benchmarks such as ARC-AGI and provides new ground for evaluating depth-recurrent reasoning models, which are typically tested on saturated Sudoku benchmarks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
compositional generalization, hypernetworks, rule learning, hierarchical reasoning models, neurosymbolic reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jQBnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a formal, controllable framework for studying knowledge transfer across small discrete worlds defined by forbidden patterns on a grid, instantiated as Sudoku-like completion problems that decompose rules into reusable components. A single depth-recurrent model is trained across many rule combinations and evaluated under four graded novelty regimes. Key findings are that description- and example-based conditioning both excel at recombining trained rules, but when rule components have never co-occurred in training, symbolic descriptions can underperform even the no-information baseline, an effect that is mitigated by increasing co-occurrence exposure or description dropout.

Critical strengths
  1. The design allows principled tests of compositional generalization, including supported and unsupported new instances.
  2. This paper proposes a strong formal grounding with an intuitive explanation.
  3. The paper is well-written and clear for readers to follow.
Critical weaknesses
  1. The experiments conducted in this paper focus on a single model family and a single grid size (6×6). It is unclear whether the scalability and architectural generality extend to larger Sudoku sizes or non-grid problems.
  2. The exploration of the example count (M=2) is limited. It is unclear that how many examples are actually needed for identifiability across different worlds is not systematically studied.
  3. There is a lack of comparison to alternative CSP solvers or LLMs with tool-use methods, which could contextualize the observed effects.
Questions for authors

See Weaknesses.

1 comment
Reviewer_jQBnOpenReview ↗

Followup Question

Thank you for the efforts made by the authors during the rebuttal. However, the applicability of the proposed framework beyond grid-based problems remains unexamined. I understand the deadline of discussion is coming, could the authors provide further discussion on how the framework could be extended to non-grid domains, whether the observed knowledge transfer patterns are expected to generalize to such settings?

Reviewer_zjUkOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper constructs Sudoku-like worlds from composable rules and separates four forms of test novelty, finding that symbolic descriptions can hurt a recurrent model on unseen rule instances even when familiar-rule recombination works well.

Critical strengths
  1. Separating new puzzles, new combinations, new co-occurrences, and unseen rule instances makes the benchmark useful for studying compositional generalization. Holding test worlds fixed while varying training exposure is a particularly helpful design.

  2. The negative-transfer finding has credible replication. In Table 3, the description model scores 34.6% on unsupported L3 instances, below the no-information model's 52.1%, while the example model reaches 65.0%. Appendix C.2 reproduces the direction across six further splits. Description dropout raises performance to 58.8%, and the controls cover sequence length, training duration, and injection mechanisms.

Critical weaknesses

I found no fatal issue with the central empirical finding. My reservations concern its explanation.

  1. The claim that capacity or symbolic format is not the bottleneck is too broad. All experiments use one recurrent architecture with fixed positional mixing. In Table 10, padding raises row-pair accuracy from 44.9 to 55.3, and composed rule vectors reach 57.9. These are meaningful differences. Fitting the training set also does not rule out an inductive bias that impedes generalization. The evidence establishes a dependence on co-occurrence for the tested model, but does not establish a general limitation of symbolic conditioning.

  2. The exposure experiments do not fully isolate co-occurrence from other changes in the training distribution. Appendix C.1 describes nested training sets while keeping 500 worlds, without explaining what is replaced as supporting or mixing worlds are introduced. Rule frequencies and puzzle difficulty may change as well. The no-information floor rises from 62.1 to 70.4 in Table 5, so some of the improvement is shared by a model that never reads the rule description.

Questions for authors

How are the exposure levels constructed while keeping 500 training worlds, and which subsets are nested? Can the low- and high-exposure endpoints be compared with rule frequencies and puzzle difficulty matched? For the architecture claim, results from the existing cross-attention variant on axis-unsupported L3 would be more informative than its L1 performance alone.

Reviewer_xgp8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper describes knowledge transfer based on different training scenarios of a depth-recurrent reasoning model related to rules and examples of Sudoku-like puzzles, which the authors describe as small discrete worlds. The tested combinations are a new puzzle, a new combination of known rules, a pairing of rules never seen together, and a rule not seen during training. All of these are presented either symbolically or with two solved examples. The baseline is the model without any rule information. The best performance is achieved with a new combination of known rules, while the worst below the baseline results from rules never seen during training. The objective is to evaluate the model's ability to extol general intelligence defined as the ability to adapt to and learn the structure of a previously unseen environment.

Critical strengths

The authors provide a detailed set of rule combinations to determine novelty for evaluation instead of just simply a held-out test set.

There is a clear baseline for the experiment that is valid and the experimental setup seems mostly rigorous even if limited.

The authors are very clear and open about limitations and potential needs to extend the proposed work.

Critical weaknesses

General intelligence claim: While it is true that benchmarks, such as ARC, rely on human conventions, including visual cues, it is not in turn automatically true that evaluation scenarios that do not rely on human conventions provide a valid test for general intelligence or intelligence beyond human performance. This is implied in the paper by stating that "this matters because general intelligence need not be human-like, and could exceed human intelligence in ways a human-anchored benchmark would fail to capture." (047-049) However, a very small discrete world of a Sudoku-like puzzles hardly provides a valid test for general intelligence and might equally fail to capture any such notion due to its limited scope. This comment naturally calls into question the provided very short account on general intelligence, but at the same time asks for more modest claims.

Limited novelty: The new contribution is the benchmark and the comparison of symbolic and example-based setups, while the other main finding that composition requires co-occurrence has already been found before.

Symbolic vs. example-based claim: the claim that the approach effectively compares a symbolic representation as opposed to an example-based one should be strongly reduced in abstract, intro, and conclusion, since the entire experiment rests on a model and training recipe as also acknowledged by the authors.

Replacing saturated benchmarks: The claim "can therefore replace already saturated fixed-game benchmarks for recurrent reasoning models" is not supported by any evidence. This is another claim I would propose reducing.

Questions for authors

Can examples of a puzzle be truly considered a form rule identity communication? In Contribution 3 it seems implied that examples represent a form of rule identity communication. Contribution 4 is equally questionable as there seems to be a too slight difference in results on presentation/communication format to warrant for a solid differentiation.

Minor comments: 024 in more systematic way => in a more systematic way 127 acronyms should be introduced 144ff What is a dial? And why use this term? 297 that each number has its floor beside it => What does this mean? 392ff Authors named in the text should come with only the year in brackets and not the full reference General: floor and dial seem to be paper-specific terms that have more common existing alternatives, e.g. baseline

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.