acceptodds
Under review as a conference paper at ICLR 2027

GeoFusion: Entanglement-Guided Fusion of Geometry-aware Representations and Spectral Priors for Code Summarization

Abstract

Code summarization aims to generate concise and accurate natural language descriptions of source code. Recent methods already exploit multi-source heterogeneous program information, yet relations with distinct structures are still placed in one Euclidean space, and their geometric properties receive limited attention. We propose GeoFusion, an entanglement-guided method that fuses geometry-aware representations with spectral priors. GeoFusion encodes abstract syntax trees (ASTs) in a Poincaré ball to preserve hierarchical syntax. Data-flow and sequential relations remain in Euclidean space, which better fits non-hierarchical dependence. To align programs with natural language, GeoFusion organizes summary semantics, source-code functionality, and line-level comment functionality in a corpus-level spectral taxonomy graph, and derives a spectral taxonomy prior. Entanglement-guided adaptive fusion then uses coupling entropy to regulate contributions across sources and within each source. Experiments on Java and Python datasets show that GeoFusion achieves the state-of-the-art performance, improving existing methods by 4.61, 4.27 and 3.02 in terms of BLEU, ROUGE-L and METEOR on Java, and by 3.78, 3.52 and 2.52 on Python.

Primary area:
generative models
Keywords:
code summarization, multi-source heterogeneous fusion, Poincaré ball, entanglement entropy

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_3Ed2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoFusion combines a Poincaré-space AST encoder with Euclidean encoders for data-flow and ordering relations. It additionally constructs a spectral taxonomy prior from associations among reference-summary categories, source-code functionality, and generated line-comment functionality. A coupling-entropy-based fusion module integrates these representations before summary generation. The paper reports improvements over the listed baselines on Java and Python, with component ablations and structural diagnostics. The main unresolved issue is how the instance-specific summary-category input to the taxonomy prior is obtained when reference summaries are unavailable at inference.

Critical strengths
  1. The relation-specific representation is a sensible modeling choice, and the paper attempts to examine it directly. Distinguishing hierarchical AST relations from dependency and ordering relations gives the architecture a coherent motivation. The mixed-, all-Euclidean-, and all-hyperbolic comparisons, together with tree-distance and subtree-separation diagnostics, are more informative than reporting only a downstream score increase. These analyses provide a useful starting point for understanding where the representation helps.

  2. The reported improvements are substantial enough to warrant attention. Relative to Hetcos, Table 1 shows BLEU gains of 4.61 points on Java and 3.78 on Python, with improvements in the other reported metrics as well. The component ablations also show that the result does not depend on a single isolated architectural change. Establishing a valid and comparable inference protocol would therefore make these results potentially consequential.

Critical weaknesses
  1. The test-time definition of the taxonomy prior is missing at a point that directly affects the validity of the main results. Section 3.2 states that the taxonomy graph is constructed from the training corpus, which is appropriate. However, the instance-specific prior also directly uses rir_i, defined as the category of the reference summary, through both a category embedding and its spectral coordinate. The paper does not specify a predictor or marginalization procedure that supplies this quantity for a new code snippet. A training-only graph does not resolve the availability of a test instance’s summary label. This is especially important because removing the summary prior causes the largest individual-view degradation in Table 2. Until the source of rir_i is explicit, the reported gains cannot be confidently interpreted as reference-free code summarization.

  2. The ablations do not adequately isolate the mechanisms used to explain the improvements. Equation (2) explicitly supplies AST depth to the radial placement, and Appendix A.2 adds hierarchy-related regularizers. Consequently, the strong depth–radius correlation is partly encouraged by construction; it is not independent evidence that curvature itself explains the downstream gain. The paper does not establish whether the Euclidean control receives comparable depth information and hierarchy supervision. Similarly, replacing EAF with concatenation or attention changes the fusion architecture, rather than isolating the contribution of coupling entropy. The current results support the full design empirically, but leave open whether simpler depth-aware encoding and an equally expressive fusion module would explain the same benefit.

Questions for authors
  1. How is rir_i obtained for validation and test samples? If it is predicted, please specify the predictor’s inputs, training data, and accuracy, and confirm whether Table 1 uses predicted or reference-derived categories. Please also clarify whether line-level comments and their labels are generated without access to the reference summaries.

  2. Does the all-Euclidean control retain the same depth information, comparable hierarchy regularization, and a matched parameter budget? For EAF, what happens when the architecture is unchanged but the entropy input and entropy-dependent gate are replaced by fixed values? These controls would distinguish the proposed geometric and entropy mechanisms from additional supervision or model capacity.

Reviewer_Lx5eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoFusion, a framework that combines Geometry-aware Representation and Spectral Taxonomy Prior for code summarization. ASTs are represented in a Poincaré ball to preserve hierarchical structures, while DFG and sequential relations are represented in Euclidean space. The framework also uses Spectral Taxonomy Prior (STP) to capture semantic relationships between source code and natural language, and Entanglement-guided Adaptive Fusion (EAF) to combine different information sources. Experiments on Java and Python datasets show improvements over existing methods in BLEU, METEOR, and ROUGE-L.

Critical strengths
  1. The paper proposes a method that combines Hyperbolic and Euclidean spaces based on the structural properties of different program relations. In particular, the proposed method captures the hierarchical structure of ASTs, which distinguishes it from existing approaches that use only Euclidean representations.
  2. The paper achieves better performance than existing methods on Java and Python datasets (Table 1). The authors also conduct ablation studies and AST structure preservation analyses to demonstrate the effectiveness of the proposed components.
Critical weaknesses
  1. There is a possibility that STP uses ground-truth information during testing. Section 3.2 uses the semantic category of the reference summary to construct STP. However, the paper does not clearly explain how this category is obtained during testing without access to the reference summary. If the category is obtained from the ground-truth summary, this could cause data leakage. This is particularly important because STP provides the largest performance improvement in the ablation study (Table 1).
  2. The fairness of the baseline comparisons and the novelty of the proposed method are not sufficiently clear. GeoFusion uses additional information from LLM-generated comments and taxonomy labels, but it is unclear whether the baselines use the same information. In addition, previous studies such as [1] have already explored hyperbolic representations of ASTs. The paper should more clearly explain the novel contribution of the proposed geometry-aware representation compared to existing methods.

[1] The Natural Geometry of Code: Hyperbolic Representation Learning for Program Reasoning, ICLR 2026

Questions for authors
  1. How is the semantic category of the reference summary used in STP obtained during inference?
  2. Could the authors confirm that no information from the ground-truth summary is used during evaluation?
  3. The paper explains that ESD operates during inference. However, Section 4.3 reports that ESD reduces the training time per epoch by approximately 70%. Could the authors explain how this is possible?
Reviewer_i4arOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoFusion for code summarization. It encodes AST relations in a Poincaré ball and dependency and ordering relations in Euclidean space. A spectral taxonomy prior combines summary-intent, code-functionality, and line-comment categories, while a coupling-entropy-based module regulates residual fusion across representations. An early-exit decoder is introduced to reduce generation cost. Experiments on Java and Python report improvements over the evaluated baselines, supported by component ablations and structural analyses.

Critical strengths

• Matching representation geometry to program structure is a reasonable and relevant motivation. The distinction between hierarchical AST relations and dependency or ordering relations leads to a concrete architecture, and the comparisons against all-Euclidean and all-hyperbolic configurations provide useful evidence for the proposed combination. • The reported improvements are consistent across two programming languages and three evaluation metrics. The component ablations and AST structural analyses go beyond an overall performance comparison and help identify which parts of the architecture warrant further investigation.

Critical weaknesses

• The spectral taxonomy prior explicitly uses reference-summary categories, but the paper does not explain how these categories are obtained at inference time. Using categories derived from test references would introduce target information, while using predicted categories requires a clearly specified and evaluated prediction procedure. This ambiguity undermines confidence in the reported improvements. • The reported 70–73% reduction in time per epoch is difficult to reconcile with the statement that training executes all decoder layers. Moreover, the ablations do not isolate spectral information from category embeddings or early exiting from intermediate supervision. These gaps limit the evidence supporting the efficiency claims and the contributions of individual mechanisms.

Questions for authors

• How are reference-summary categories obtained for unseen test samples? Are they predicted solely from source code, and how does using predicted rather than gold categories affect performance? • What does the reported time per epoch include? How does early exiting achieve a 70–73% reduction when training executes all decoder layers? • How much improvement comes from spectral coordinates beyond category embeddings? Have you evaluated early exiting separately from intermediate supervision?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.