acceptodds
Under review as a conference paper at ICLR 2027

From Performance to Capability: A Three-Level Framework for Behavioral LLM Capability Science with Large-Scale Empirical Study

Abstract

Behavioral LLM capability science asks how evidence from observable behavior can justify increasingly strong claims about what an LLM system can do and how its behavior is organized. We propose a three-level framework that distinguishes (1) Observable Performance, (2) Conditional Response and Capability Validity, and (3) Fundamental Organization, and makes the evidential transitions between them explicit and testable. We evaluate this framework across 34 heterogeneous released source collections spanning reasoning, feedback, memory, prompting, planning, coding, retrieval, and related settings, comprising over one million task-level responses. Conditional Response emerges as a substantive empirical object: response profiles reproduce across compatible roots (rank correlation 0.586), 73/311 eligible groups exhibit held-root nonconstant regularities, and multiple designs show nonadditive effects. Stronger capability claims require connected evidence involving reliability, transfer, convergence across operationalizations and sources, specificity, discrimination, measurement quality, controls, external prediction, and compatible common support. Four of nine candidate constructs obtain local reliability and task-transfer support, and one obtains cross-operationalization support, but none completes connected validity because key links in the released evidence topology remain missing or conflicting. Human-readable semantic and conditional structure is predictively useful on held-out data, yet predictive or low-dimensional structure alone does not establish Fundamental Organization. The resulting evidence frontier distinguishes observed performance, established response organization, supported capability evidence, and unresolved deeper organization, providing a concrete foundation for cumulative LLM capability science.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, large language model evaluation, large language model capability evaluation, large language model capabilities, large language model capability inference, large language model benchmarking, large language model behavioral evaluation, large language model evaluation methodology

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

This paper proposes a three-level framework for studying LLM capabilities, distinguishing observable performance, conditional responses and capability validity, and fundamental organization. The authors analyze 34 previously released datasets, covering more than one million task-level responses, to examine the proposed framework. The results show reproducible patterns in conditional responses, but none of the nine candidate capabilities meets the full validity requirements. The study also finds useful predictive structure, although the evidence is insufficient to establish fundamental organization.

Critical strengths
  1. The paper addresses an important problem: how to distinguish benchmark performance from evidence of underlying model capabilities. The proposed framework encourages more careful interpretation of evaluation results.

  2. The authors analyze a large collection of existing datasets and apply several statistical tests. Their distinction between unsupported claims and insufficient evidence is useful, and they generally acknowledge the limitations of their findings.

Critical weaknesses
  1. The main theoretical contribution is not sufficiently clear. The distinction between performance and capability, along with the need for construct validity, has been discussed extensively in previous work. The paper combines these ideas into a three-level framework, but does not clearly show what new understanding this framework provides beyond existing approaches.

  2. The empirical analyses do not adequately validate the proposed framework. None of the nine candidate capabilities satisfies the full validity criteria, largely because the existing datasets lack the necessary connections. The paper also provides no complete real-world example showing how the three levels work together. As a result, the analyses mainly document limitations of the available data rather than demonstrate the effectiveness of the framework.

Questions for authors
  1. What does the three-level framework add to existing approaches to behavioral evaluation and construct validity? Could the authors give a concrete example of a conclusion that becomes possible under this framework but not under previous approaches?
  2. Given that none of the nine candidate capabilities meets the full validity requirements, how can the authors demonstrate that the framework is practically useful? Is there a real dataset that would allow a more complete validation?
  3. How were the nine candidate capabilities selected, and how sensitive are the conclusions to these choices and the proposed validity criteria?
Reviewer_96axOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a three-level conceptual framework to formalize behavioral LLM capability science: Level 1 (Observable Performance YsrcY_{src}), Level 2 (Conditional Response Rsr(c,c0)=Ysrc−Ysrc0R_{sr}(c, c_0) = Y_{src} - Y_{src0} and a multi-criteria "connected validity" contract), and Level 3 (Fundamental Organization evaluated across five structural questions). Applying this framework to a pooled secondary dataset of 34 released benchmark collections (yielding 10.4M raw observations, 1.0M task-level responses, and 6,798 response cells across 16 core sources), the authors report that:

  • Conditional response profiles replicate across task-root splits (median rank correlation 0.5860.586)
  • 73 of 311 candidate law groups display held-root nonconstant relations, with 6 supported interaction tests
  • Construct-level semantic indicators moderately outperform a masked low-rank comparator in predicting out-of-distribution performance (ΔMSE=−0.0115\Delta\text{MSE} = -0.0115)
Critical strengths
  1. The framework strictly differentiates between genuine model behavioral failures and measurement artifacts (e.g., distinguishing VALID, VALID_ZERO, POLICY_REJECTION, and ENVIRONMENT_FAILURE), establishing a disciplined taxonomy for secondary benchmark meta-analysis.
  2. Qualification thresholds and promotion rules are tuned on separate synthetic worlds, demonstrating zero false-positive promotions (0/1000/100) under null conditions.
  3. The paper avoids over-claiming by separating "not identifiable" (data-topology limits) from empirical refutation, and explicitly documents full interaction and reversal universes rather than reporting only significant subsets.
  4. Employs root-clustered uncertainty, global Benjamini-Hochberg/Yekutieli adjustments, coherent empirical-null replays, and outcome-unconditioned split-half reliability calculations.
  5. The analysis provides a valuable diagnosis of fragmented benchmark infrastructure, pinpointing how the absence of shared scaffolds, common versioning, and repeat IDs prevents cross-study synthesis.
Critical weaknesses
  1. The conceptual framework largely renames established psychometric and evaluation principles without introducing formal identification theorems:
  • Level 2's validity criteria map directly onto classical construct validity (Cronbach & Meehl 1955), Campbell & Fiske's Multitrait-Multimethod (MTMM) matrix, Messick's unified validity, and Kane's argument-based validation.
  • Recent literature has already proposed nomological networks and capability-centered frameworks for LLMs (Freiesleben 2026; Jo & Wilson 2026; Bean et al. 2025; Burden et al. 2023).
  • "Conditional response" is a standard within-item difference score. Difference scores suffer from well-documented psychometric pitfalls, notably attenuated reliability and mathematical coupling to baseline performance (Cronbach & Furby 1970), which the paper notes only descriptively.
  1. The headline conclusion that zero candidate constructs complete connected validity (8/9 NI) is an artifact of the authors' inclusion criteria rather than an empirical finding about models. Requiring exact model-version and exact harness/scaffold identity across independent datasets ensures zero cross-source overlap by construction, because external benchmarks intentionally deploy task-specific wrappers. In standard MTMM logic, scaffolds represent method facets; requiring scaffold invariance precludes assessing convergent validity across heterogeneous methods. When model versions are matched across varying scaffolds, substantial cross-source overlap exists in the pooled data (e.g., GLM-5.1, MiniMax-M2.7, GPT-4o).
  2. The headline figure of 10.4M observations is misleadingly dominated by a single source (DOVE-Lite accounts for 86.4% of rows but reduces to 26 response cells across 4 models). Across most candidate panels, the effective sample size is only 4 to 17 models, limiting the statistical power of broad capability claims.
  3. The recovery pipeline achieves only 29% recovery (14/48) on synthetic ground-truth worlds even under strong signal, scoring 0% across modular and global families. The failure to establish Level-3 organizational structure reflects the low statistical power of the recovery instrument and sparse data topology, not the absence of latent structure in LLMs.
  4. The main text relies heavily on undefined coined terminology ("evidence topology," "canonical task roots," "structured-unpaired," "fail closed") and is not self-contained. The operational definitions of the nine candidates and their data mappings are deferred entirely to the 46-page appendix.
Questions for authors
  1. If you relax the exact scaffold identity requirement to evaluate exact model versions across different scaffolds (treating scaffold as a method facet in an MTMM framework), how many cross-source links become estimable, and do any candidate validity states change?
  2. Are the six systems shared between semantic adaptation and reasoning integrity identical to C13 (SEAL)? What is the discriminant correlation when evaluated strictly on mutually disjoint task rows?
  3. What is the split-half rank correlation of raw Level-1 performance across the same 1,009 root groups? How much of the 0.5860.586 conditional response correlation remains after controlling for baseline performance and ceiling constraints?
  4. How many of the 73 nonconstant laws and 6 interaction effects remain supported if proxy-scored runs from C17 are excluded?
  5. When the synthetic calibration pipeline is evaluated on a topology matched to the real pooled dataset (zero cross-source identical bridges), what is the empirical promotion power?
Reviewer_J3GWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how observable LLM performance can support claims about underlying capabilities. It proposes a three-level framework: (a) Observable Performance; (b) Conditional Response and Capability Validity; and (c) Fundamental Organization. The authors analyze 34 existing datasets (16 in the core analysis), covering over 1 million task-level responses across reasoning, coding, memory, planning. The paper provides the following main findings: (a) conditional response profiles show moderate reproducibility across task splits (Spearman correlation of 0.586); (b) of 9 candidate capabilities, 4 show local reliability and transfer support, but none meets all validity requirements; (c) semantic features outperform a low-rank representation on held-out task prediction, while conditional-response information improves prediction beyond overall model strength; (d) the available evidence does not establish a stable, transferable organizational structure.

Critical strengths
  • The papers analysis spans across extensive settings. It uses 34 existing research datasets spanning reasoning, coding, feedback, memory, planning. It also conduct several statistical analyses, including response reproducibility, held-out prediction, synthetic calibration, and robustness checks.
  • The proposed framework distinguishes observable performance from conditional responses, validated capabilities, and underlying organization. The authors find reproducible behavioral response patterns and demonstrate that local predictive or behavioral evidence does not necessarily establish a validated capability. This provides useful insights beyond conventional benchmark accuracy.
Critical weaknesses
  • The paper does not demonstrate successful end-to-end capability validation on real-world data, making it difficult to assess the framework's practical effectiveness. The paper proposes an ambitious three-level framework, but none of the 9 candidate capabilities achieves full validation. Many required relationships cannot be tested due to limitations of the existing datasets. A successful demonstration on real-world data would help establish the framework's practical effectiveness.
  • The novelty and generalization of the work is limited. The distinction between performance and capability has been explored in prior works. The paper mainly unifies existing ideas into a framework without clearly demonstrating its advantages over prior approaches, making the novelty limited. Further, despite analyzing 34 datasets, the positive findings are concentrated in smaller subsets. For example, 54 of the 73 supported non-constant relationships come from a single source. This raises concerns about generalizability.

Overall, the framework is useful for organizing evidence, but its key claims remain insufficiently validated on real-world data.

Questions for authors
  • Can the authors demonstrate the complete framework in at least one real-world setting? This would help establish its practical usefulness beyond identifying evidence gaps.
  • How well do the Conditional Response findings generalize across datasets and model families? Can you demonstrate that the results are not primarily driven by a few datasets?
  • How does the framework differ from prior work (https://arxiv.org/pdf/2511.04703, https://arxiv.org/pdf/2509.19590, https://arxiv.org/pdf/2503.06378)? What additional conclusions does your framework enable?
  • How sensitive are the findings to the choice of candidate capabilities, their operationalizations, and validity criteria? How will the alternatives change the findings?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.