acceptodds
Under review as a conference paper at ICLR 2027

TableLoop: Tabularizing Multimodal Data via Rubric Evolution

Abstract

Many prediction tasks depend on information distributed across text, images, and structured measurements. We introduce TableLoop, a flexible framework that connects language and vision-language models with tabular predictors through learned natural-language rubrics, closing the loop between what a model measures and how useful those measurements are for downstream prediction. A frozen scoring model applies each rubric to produce explicit numeric features, while a tabular predictor's validation performance and errors guide iterative revisions to the rubric. The resulting interpretable features are combined with structured measurements for final prediction. TableLoop is modular: it requires no language-model tuning and supports interchangeable scoring models, tabular predictors, and evolution methods. Across nine tasks spanning clinical risk, persuasion outcomes, essay quality, peer-review decisions, skin-lesion malignancy, and fungal species identity, TableLoop outperforms both structured-data-only prediction and calibrated direct prediction by a larger language or vision-language model in 26 of 27 evaluated configurations. Notably, TableLoop using a smaller open-weight scorer (Qwen3-32B/Qwen3-VL-32B) can outperform direct prediction with a frontier model (Claude Sonnet 4.5), showing that a smaller model can support stronger predictions when used to extract task-specific features rather than predict the target directly. Beyond prediction, TableLoop produces reusable, interpretable feature tables grounded in real observations, offering a potential source of training and adaptation data for tabular foundation models and an inspectable interface for multimodal-to-table learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal learning, Representation learning, Large language models, Tabular learning, Rubric evolution

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_875dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

TableLoop converts unstructured inputs into numeric features using learned rubrics, then combines these features with structured measurements for a tabular predictor. Rubrics are revised using validation performance and instance-level errors. The paper also introduces SH-merge, a successive-halving method that merges complementary rubrics. It is evaluated on nine tasks across text and image inputs, including classification and regression.

Critical strengths

Broad and consistent empirical validation across nine tasks, multiple modalities, and different evolution methods.

The learned rubrics produce inspectable, reusable feature tables, which is a practical advantage over end-to-end LLM prediction.

Thorough ablations covering initialization, evolution method, instance-level feedback, metadata integration, predictor choice, and training subsampling.

Critical weaknesses

The direct LLM/VLM baseline is zero-shot with affine calibration, not a fine-tuned or otherwise stronger baseline. This limits how much of the gain is attributable to TableLoop versus the choice of baseline.

Rubric evolution is computationally expensive. The paper acknowledges this and shows training subsampling helps, but the full search budget remains heavy.

The retrospective truncation analysis evaluates the test set at every iteration. The main experiments use a fixed budget and evaluate the test set once, so the concern is limited, but the analysis should be clarified.

Questions for authors

How does TableLoop compare to fine-tuned LLMs or strong multi-modal AutoML baselines?

What is the total computational cost of rubric search across the nine tasks?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 875d (1/2)

We thank the reviewer for acknowledging the practical value of TableLoop’s inspectable, reusable feature tables, the breadth of our evaluation across tasks and modalities, and the thoroughness of our ablation studies. We reply to each of the points raised in the review below. We have also uploaded a revised version of the manuscript highlighting the main changes in blue color.

Weakness 1 and Question 1: How does TableLoop compare to fine-tuned LLMs or strong multi-modal AutoML baselines?

Response: We thank the reviewer for the valuable question. We added a fine-tuned direct predictor and a supervised multimodal AutoML baseline in Appendix J.

LoRA fine-tuning. On Persuasion, we fine-tuned Qwen3-32B to predict donations using the same dialogues, metadata, and data splits as TableLoop. The fine-tuned model achieved a test ROC-AUC of 0.7930, while TableLoop (OpenEvolve) achieved 0.8393 with a frozen scorer (Appendix J.2).

Supervised multimodal AutoML. We evaluated AutoGluon-Multimodal (AutoMM) on seven tasks using the same source inputs, metadata, splits, and metrics as TableLoop (Appendix J.1). We used the best-quality preset with augmentation and also evaluated a long-context ModernBERT variant for text. Both configurations are reported below where applicable.

TaskTest metricAutoMM (Default)AutoMM (Long-context)TableLoop (OpenEvolve)
PersuasionROC-AUC ↑0.70370.71140.8393
ELLIPSEMAE ↓0.32700.32070.3334
MIMIC-MVPR-AUC ↑0.51590.52380.5218
ICLR (RL)ROC-AUC ↑0.56840.58620.7014
ICLR (Graphs)ROC-AUC ↑0.65090.66130.6946
PAD-UFES-20ROC-AUC ↑0.9750—0.9734
Danish FungiROC-AUC ↑0.9600—0.9310

The results vary by task. TableLoop outperforms both AutoMM configurations on Persuasion and both ICLR tasks, with the largest gains over the stronger AutoMM configuration on Persuasion and ICLR (RL), at 0.1279 and 0.1152 ROC-AUC, respectively. The methods perform similarly on MIMIC-MV and PAD-UFES-20, while AutoMM performs better on ELLIPSE and Danish Fungi. This pattern is consistent with our hypothesis that rubric-based scoring is particularly useful when prediction involves complex reasoning over text.

AuthorsOpenReview ↗

Response to Reviewer 875d (2/2)

Weakness 2 and Question 2: What is the total computational cost of rubric search across the nine tasks?

Response: Thank you for raising this point. We added a runtime and computational cost analysis in Appendix L (Tables 23–25), reporting per-example GPU time, scoring throughput, and GPU-hour breakdowns for an offline candidate evaluation and a held-out evaluation with one selected rubric. Importantly, Appendix L reports measured GPU-hours for complete training- and validation-set scoring and TabPFN evaluation per candidate, while Appendix I.2 specifies the search budgets for all three evolution methods. Together, these measurements quantify the computational scale of rubric search rather than relying solely on theoretical complexity or per-token estimates. We agree that offline search is relatively costly because it requires repeatedly scoring training and validation examples. This upfront cost is part of learning an interpretable representation in which each feature is defined by an explicit scoring criterion that can be inspected and reused. In the online stage, TableLoop applies the selected rubrics to new inputs and reuses the fitted tabular predictors, which is much less costly than running the full rubric search.

Several strategies can further reduce the cost of both search and prediction. The number of scoring calls during search can be reduced by evaluating fewer candidates through early stopping or by scoring fewer training examples per candidate. On ELLIPSE, using just 10% of the training examples per candidate while retaining the full validation set yielded comparable test MAE to full-data search (0.3323 vs. 0.3334, Appendix F). Fewer scoring repetitions can further reduce the number of calls, while parallel scoring can shorten wall-clock time in both stages. Most of the Qwen scorers’ GPU time goes to autoregressive decoding, and our benchmark shows that JEV System 1 (9B) achieves a roughly 5× speedup over Qwen3-8B in per-example GPU time on the same rubrics (Table 23). A promising direction for future work is to test whether JEV can reduce TableLoop’s end-to-end runtime without sacrificing predictive performance.

Weakness 3: The retrospective truncation analysis evaluates the test set at every iteration. The main experiments use a fixed budget and evaluate the test set once, so the concern is limited, but the analysis should be clarified.

Response: We thank the reviewer for this question. As described in Section 5.1, the main OpenEvolve experiments use a fixed budget of 50 iterations. Validation feedback guides rubric revisions and selection, and we evaluate the final ensemble on the test set only once. For Figure 3, after the runs were complete, we reconstructed the top-three ensemble at each checkpoint using validation performance alone and evaluated it on the test set. These retrospective evaluations show how held-out performance changes with the search budget and did not affect how we revised the rubrics, which rubrics we selected for the ensemble, or when we stopped the search.

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

TableLoop combines unstructured text or images with structured measurements through learned natural-language rubrics. A frozen language or vision-language model converts unstructured inputs into numerical rubric scores. A tabular predictor combines these features with metadata, and its validation performance and errors guide rubric revisions. The paper evaluates OpenEvolve, GEPA, and a proposed successive-halving-with-merge method across nine tasks from six data sources. It reports improvements over both metadata-only and calibrated direct-prediction baselines in 26 of 27 main configurations. It also compares against a reimplementation of FEST on seven binary tasks. Additional experiments examine feedback, initialization, rubric selection, metadata, and downstream predictors.

Critical strengths
  1. A workable framework with cross-task evidence. The implementation connects frozen language models with supervised predictors across text and image tasks, using several search methods and downstream predictors. OpenEvolve and GEPA outperform the reimplemented FEST baseline on all seven binary tasks under matched data splits, metadata, and final predictors, providing a relevant comparison beyond direct prediction.
  2. Useful evidence for predictive-feedback optimization. Figure 3 shows that rubric evolution generally improves held-out performance. Table 1 isolates instance-level feedback while retaining aggregate validation metrics, with substantial effects on Persuasion and MIMIC-MV. Appendix D also supports direct rubric evolution over instruction evolution on two tasks.
Critical weaknesses
  1. The comparison does not isolate the advantage of tabularization. Inspectability and compatibility with metadata provide a reasonable motivation for rubric features, but the claimed advantage over direct prediction needs stronger attribution. The main direct-prediction baseline receives both unstructured inputs and metadata, with validation-based calibration; TableLoop additionally benefits from supervised prediction, iterative labeled-error feedback, rubric search, and ensemble selection. Appendix B's few-shot comparisons do not match this adaptation process. The results support the complete pipeline's advantage over the tested direct predictors, but leave unclear what converting inputs into rubric scores and using a tabular decision model contributes beyond the additional task adaptation.
  2. Competitiveness against established multimodal prediction approaches remains untested. The FEST comparison supports TableLoop relative to one related feature-engineering approach, but the evaluation lacks representative supervised multimodal predictors such as AutoMM under matched protocols. This leaves the practical significance of the reported gains difficult to assess. External benchmark results cannot substitute for these comparisons. For example, Appendix H.1 defines a five-species Danish Fungi task with a new split and observation-level macro ROC-AUC. The protocol is disclosed, but the rationale for this subset and its representativeness are insufficiently explained. Strong comparisons within the chosen tasks are needed to establish the framework's contribution.
Questions for authors
  1. What evidence separates the benefit of tabularization and tabular decision-making from that of iterative task adaptation? One informative control would use a strong LLM/agent for decisions with access to the same information, training examples, rubric knowledge and tools, and a comparable feedback and optimization budget.
  2. How does TableLoop compare with representative supervised multimodal predictors under the same protocol? Please also explain the choice and representativeness of the five-species Danish Fungi subset.
2 comments
AuthorsOpenReview ↗

Response to Reviewer ChLB (1/2)

We appreciate the reviewer’s detailed comments and their recognition of TableLoop’s flexibility across text and image tasks, its gains over FEST under matched conditions, and the evidence supporting feedback-guided rubric evolution. Below, we address all of the points raised.

Weakness 1 and Question 1: What evidence separates the benefit of tabularization and tabular decision-making from that of iterative task adaptation?

Response: Thank you for the insightful question. We added two experiments in Appendix K to help separate the contribution of rubric-based tabular prediction from that of iterative task adaptation.

Prompt-optimized direct prediction. We compare the GEPA version of TableLoop with GEPA-optimized direct prediction under the same search budget and a matched validation feedback protocol, using the same data splits, metadata features, and top-three ensembling (Appendix K.2). Both searches use Claude Sonnet 4.5 to propose revisions and the same task-specific Qwen model to produce rubric scores or direct predictions. Both direct-prediction variants receive the input and its metadata. The 0-shot prompt contains no demonstrations, while the 5-shot prompt includes five training examples retrieved by embedding similarity, each with its label and metadata. The direct-prediction columns below report test results obtained using Claude Sonnet 4.5 with the selected prompts (Table 22). TableLoop outperforms both prompt-optimized direct predictors on all three tasks. This supports the use of rubric features with a tabular predictor even when direct prediction also receives iterative validation feedback under the same search budget.

TaskTest metricGEPA direct (0-shot)GEPA direct (5-shot)TableLoop (GEPA)
PersuasionROC-AUC ↑0.79980.80250.8275
ELLIPSEMAE ↓0.37970.36190.3358
PAD-UFES-20ROC-AUC ↑0.87340.93370.9733

Rubric generation with fixed dimensionality. To examine what iterative evolution contributes beyond rubric-based tabularization and finding a suitable rubric size, we compare full TableLoop with zero-shot rubric generation and a single round of TableLoop (Appendix K.1, Table 21). Each baseline generates three rubrics per task, with dimension counts matched one-to-one to those of the top three rubrics from the full OpenEvolve run. We use the same scorer, metadata, TabPFN predictor, and ensembling procedure across settings. Full TableLoop outperforms both baselines on six of seven tasks and ties the one-round variant on Danish Fungi. These results suggest that iterative refinement can improve predictive performance even when the baselines use the same tabularization pipeline and inherit the rubric sizes found by full evolution.

Together, the two ablations provide complementary controls: fixed-dimensional rubric generation holds the tabular prediction pipeline fixed while varying evolutionary refinement, whereas GEPA-optimized direct prediction holds the search procedure and evaluation budget fixed while comparing direct prediction against rubric-based tabular prediction. These experiments provide strong evidence that both evolutionary optimization and rubric-based tabular prediction are important contributors to TableLoop’s performance.

AuthorsOpenReview ↗

Response to Reviewer ChLB (2/2)

Weakness 2 and Question 2: How does TableLoop compare with representative supervised multimodal predictors under the same protocol? Please also explain the choice and representativeness of the five-species Danish Fungi subset.

Response: We thank the reviewer for the question. We evaluated AutoMM on seven of our tasks, using the same source inputs, metadata, data splits, and evaluation metrics as TableLoop (Appendix J.1, Table 20). We use AutoMM’s best-quality preset with augmentation and also test a long-context ModernBERT variant for text. The table below shows the tested AutoMM results for each task.

TaskTest metricAutoMM (Default)AutoMM (Long-context)TableLoop (OpenEvolve)
PersuasionROC-AUC ↑0.70370.71140.8393
ELLIPSEMAE ↓0.32700.32070.3334
MIMIC-MVPR-AUC ↑0.51590.52380.5218
ICLR (RL)ROC-AUC ↑0.56840.58620.7014
ICLR (Graphs)ROC-AUC ↑0.65090.66130.6946
PAD-UFES-20ROC-AUC ↑0.9750—0.9734
Danish FungiROC-AUC ↑0.9600—0.9310

TableLoop leads on Persuasion and both ICLR tasks. The methods obtain similar scores on MIMIC-MV and PAD-UFES-20, while AutoMM leads on ELLIPSE and Danish Fungi. The large gains on Persuasion and ICLR acceptance prediction also appear in the frozen-embedding comparison (Section 5.5). We hypothesize that rubric-based scoring yields larger gains when a task calls for more complex reasoning over text.

For the Danish Fungi task, we chose five species from the genus Russula with overlapping visual traits, including similar cap colors and pale gills. This gives us a fine-grained, five-class image classification task in which all methods receive field photographs and observation metadata (Appendix I.1). The subset is designed to test discrimination among visually similar species and is not intended to represent the taxonomic diversity of the full Danish Fungi dataset.

Reviewer_4TbBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TableLoop, a framework that uses LLM-driven rubric evolution to extract interpretable numerical features from text and images for downstream tabular prediction. Experiments across nine tasks demonstrate improvements over metadata-only prediction and direct LLM/VLM prediction

Critical strengths
  1. Interesting framework for task-driven, interpretable feature discovery without LLM fine-tuning.

  2. Consistent improvements across diverse datasets and prediction tasks.

  3. Useful ablations on evolution strategies, validation feedback, and rubric initialization.

Critical weaknesses
  1. Insufficient baselines. The paper lacks comparisons with frozen LLM/VLM embeddings + TabPFN and fixed or randomly generated rubrics with matched feature dimensionality. It is unclear how much improvement comes specifically from rubric evolution.

  2. Limited interpretability evaluation. Although interpretability is a central claim, the paper does not sufficiently examine whether evolved rubric scores faithfully represent their intended concepts.

  3. Potential validation overfitting. Repeatedly using validation labels and instance-level errors to evolve rubrics may lead to overfitting. Multi-seed and cross-split stability experiments are needed.

  4. Computational cost. Repeated LLM/VLM scoring may be expensive, but comprehensive runtime and cost comparisons are missing.

Questions for authors
  1. How do rubric dimensions and definitions change throughout evolution? It would be good if concrete examples are provided.

  2. How does TableLoop compare with frozen embeddings followed by the same TabPFN predictor?

  3. Are the evolved rubrics stable across different seeds, data splits, and scoring models?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4TbB (1/2)

We thank the reviewer for the thoughtful feedback and for recognizing the value of interpretable feature discovery and our ablation studies. We reply to each of the points raised in the review below and we have also uploaded a revised version of the manuscript highlighting the main changes in blue.

Weakness 1 and Question 2: How does TableLoop compare with frozen embeddings followed by TabPFN and with rubrics of matched dimensionality?

Response: We thank the reviewer for the valuable question. We added both comparisons. Section 5.5 (Table 4) compares TableLoop with frozen Qwen3 text and vision-language embeddings using the same TabPFN predictor, data splits, and metrics. We evaluate the original and PCA-reduced embeddings, each with and without metadata. TableLoop outperforms all four embedding variants on six of seven tasks, including all five text tasks. The table below shows the best embedding result for each task.

TaskTest metricBest embeddingTableLoop
PersuasionROC-AUC ↑0.77030.8393
ELLIPSEMAE ↓0.37430.3334
MIMIC-MVPR-AUC ↑0.47910.5218
ICLR (RL)ROC-AUC ↑0.59200.7014
ICLR (Graphs)ROC-AUC ↑0.65460.6946
PAD-UFES-20ROC-AUC ↑0.97430.9734
Danish FungiROC-AUC ↑0.91980.9310

To examine what evolution contributes beyond finding a suitable rubric size, we compare full TableLoop with zero-shot rubric generation and a single round of TableLoop (Appendix K.1). Each baseline generates three rubrics per task, with dimension counts matched one-to-one to those of the top three rubrics from the full TableLoop run with OpenEvolve. Both baselines use the same scorer, metadata, TabPFN predictor, and ensembling procedure as full TableLoop. Full TableLoop outperforms both baselines on six of seven tasks and ties the one-round variant on Danish Fungi. These results suggest that iterative rubric evolution improves predictive performance beyond what can be achieved by matching rubric dimensionality alone.

Across both comparisons, TableLoop’s gains are particularly pronounced on Persuasion and ICLR acceptance prediction. We hypothesize that task-adapted rubrics offer greater benefits on tasks that require more complex reasoning over text, where explicit criteria can guide the scorer to assess evidence relevant to the prediction target.

Weakness 2 and Question 1: How do rubric dimensions and definitions change during evolution, and do the scores reflect their intended concepts?

Response: We thank the reviewer for bringing this up. We added concrete examples in Appendix C showing how rubric dimensions and scoring instructions change over evolution. On MIMIC-MV, the rubric moves from a single estimate of ventilation risk to separate assessments of respiratory distress and radiographic abnormalities, with later instructions specifying how to assess findings such as the physiological impact of atelectasis. On ICLR (Graphs), it moves from an overall paper-quality score to graph-specific criteria that specify what evidence earns a high score, such as a proof of a new expressiveness result rather than a reference to an established hierarchy.

Our interpretability claim concerns the explicit, human-readable definitions of the extracted features and their inspectable instance-level scores, rather than a claim that the evolved dimensions have been independently validated against human concept annotations. The examples in Appendix C illustrate this transparency, while evaluating semantic faithfulness remains an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer 4TbB (2/2)

Weakness 3 and Question 3: Does repeated validation feedback lead to overfitting? Are the evolved rubrics stable across seeds, data splits, and scoring models?

Response: We thank the reviewer for this comment. To check whether the gains carry over to held-out data, we retrospectively evaluated the validation-selected ensemble at each checkpoint of the fixed-budget runs. Held-out test performance generally improved over the course of evolution (Figure 3).

We also ran some checks on predictive stability. Rerunning OpenEvolve on Danish Fungi yielded a test macro ROC-AUC of 0.9311, compared with 0.9310 in the original run. On Persuasion, OpenEvolve achieved a test ROC-AUC of 0.8308 with Qwen3-8B as the scorer (vs. 0.8393 with Qwen3-32B), still above the direct-prediction and frozen-embedding baselines. Also, our main ICLR results report mean test ROC-AUC across three stratified validation/test splits, with the training set held fixed (Appendix I.1). Given the time available for this revision, we did not conduct a broader study of whether the learned rubric criteria remain consistent across seeds, data splits, and scoring models. We agree that this is an important future direction and note it in the Discussion.

Weakness 4: Computational cost. Repeated LLM/VLM scoring may be expensive, but comprehensive runtime and cost comparisons are missing.

Response: Thank you for raising this point. We added a runtime and computational cost analysis (Appendix L, Tables 23–25) that reports per-example GPU time, scoring throughput, and GPU-hours for an offline candidate evaluation and a held-out evaluation with a selected rubric. During offline rubric search, TableLoop repeatedly scores training and validation examples as the rubrics evolve. This upfront cost is part of learning an interpretable representation, where each feature corresponds to an explicit scoring criterion. During the online stage, TableLoop applies the rubrics selected offline to new inputs and reuses the fitted tabular predictors, making prediction much less costly than rubric search.

We also discuss several ways to make rubric search and prediction more efficient. Early stopping, training-set subsampling, and fewer scoring repetitions can reduce the number of scoring calls during search, while parallel scoring can shorten wall-clock time in both stages. On ELLIPSE, using 10% of the training examples per candidate while retaining the full validation set yielded comparable test MAE (Appendix F). Finally, most of the Qwen scorers’ GPU time goes to autoregressive decoding, while JEV System 1 (9B) achieves a roughly 5× speedup over Qwen3-8B in per-example GPU time on the same rubrics (Table 23). A promising direction for future work is to test whether JEV can reduce TableLoop’s end-to-end runtime without sacrificing predictive performance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Main Response to All Reviewers

We thank the reviewers for their thoughtful and constructive feedback. We are encouraged by their positive assessment of TableLoop’s central idea: using predictive feedback to learn interpretable, reusable features from text and images without fine-tuning the scoring model. Our experiments across nine tasks show how this approach works in different settings, and the ablations help explain how rubric evolution and validation feedback contribute to its performance.

In response to the reviewers’ suggestions, we revised the manuscript and highlighted the main changes in blue. The main revisions are summarized below:

  • How does TableLoop compare with frozen embeddings? (4TbB W1/Q2) We added a comparison with frozen LLM and VLM embeddings using the same TabPFN predictor (Section 5.5). Task-adapted rubric features generally outperform the frozen embeddings.
  • How do rubrics change during evolution? (4TbB Q1) We added examples from MIMIC-MV and ICLR (Graphs) showing how evolution turns broad initial criteria into more specific, task-relevant scoring dimensions (Appendix C).
  • How does TableLoop compare with supervised multimodal and fine-tuned baselines? (ChLB W2/Q2, 875d W1/Q1) We added comparisons with AutoMM and a LoRA-fine-tuned Qwen3-32B direct predictor (Appendix J). TableLoop outperforms the fine-tuned direct predictor on Persuasion, while the results against AutoMM are more mixed.
  • How do tabularization and evolutionary optimization contribute to TableLoop’s gains? (4TbB W1, ChLB W1/Q1) We added two comparisons in Appendix K: full rubric evolution versus zero-shot and one-round rubric generation at matched feature dimensionalities, and TableLoop versus direct prediction with GEPA prompt optimization under the same evolution budget. Full evolution generally improves on the matched-dimensional rubric baselines, and TableLoop consistently outperforms the prompt-optimized direct predictor in the tested settings.
  • What is the computational cost of rubric search and prediction? (4TbB W4, 875d W2/Q2) We added runtime and GPU cost measurements for offline candidate evaluation and held-out evaluation (Appendix L).

We also made minor numerical corrections to Table 5 in Appendix A.

Among these additions, the frozen-embedding comparison addresses whether task-adapted rubric scores offer an advantage over general-purpose representations when both are used with the same tabular predictor. We keep the tabular model (TabPFN), the data splits, and the evaluation metrics fixed. The table below compares TableLoop with the best-performing frozen embedding configuration on each task. The full results are reported in Section 5.5 (Table 4).

TaskTest metricBest embeddingTableLoop
PersuasionROC-AUC ↑0.77030.8393
ELLIPSEMAE ↓0.37430.3334
MIMIC-MVPR-AUC ↑0.47910.5218
ICLR (RL)ROC-AUC ↑0.59200.7014
ICLR (Graphs)ROC-AUC ↑0.65460.6946
PAD-UFES-20ROC-AUC ↑0.97430.9734
Danish FungiROC-AUC ↑0.91980.9310

TableLoop outperforms the frozen embedding baselines on six of seven tasks, including all five text tasks. The largest gains are on ICLR (RL) and Persuasion, with ROC-AUC improvements of 0.1094 and 0.0690, respectively. These results show that task-adapted rubric features can improve held-out performance over general-purpose embeddings with the same downstream predictor.

The new ICLR experiments in this revision use the recently released PaperScorer-Text dataset as an alternative source of paper text (Appendix I.4). In the final manuscript, we plan to update the remaining ICLR experiments to use PaperScorer-Text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.