A Correct Answer Is Not a Complete Model Profile

TL;DR for operators

A model’s correctness score tells you whether its answers passed a test, but discards information about how it responds. BULLPEN shows that retaining answer text substantially improves prediction of behavioral, stylistic, and provenance characteristics across 260 open-weight models. Its answer-only representation achieves 0.48 normalized predictive skill on 20 new tasks, compared with 0.35 for the strongest prior encoder.

The operational opportunity is cheaper, broader model screening. In a separate evaluation involving 104 models, 400 selected questions preserve 91% of measured full-pool skill for BULLPEN’s fused representation and approximately all of it for its answer-only representation, using roughly 1.3% of full-pool inference compute.

The qualification is consequential: correctness-based representations still perform better on established competence tasks, and combining all three information channels does not demonstrate consistent superiority over the strongest individual channel. Model-selection teams should treat competence and behavioral profiling as distinct assessments, not assume that one embedding can replace both.

Two models can pass the same tests and behave differently

Consider a team selecting an open-weight language model for customer-facing assistance. Two candidates achieve similar benchmark accuracy. One answers concisely; the other regularly produces long explanations, hedges, or refuses requests that the first model completes.

A correctness-based evaluation can give these models nearly identical profiles. It records whether an answer satisfies the grading criterion but usually discards the wording that produced the judgment.

That discarded information may influence deployment suitability. Response length affects interaction cost and user experience. Refusal patterns affect task completion. Stylistic tendencies can influence preference ratings even when factual correctness is unchanged.

Russell Ro, Qiming Cui, and Reza Abbasi-Asl investigate this gap in From the Zoo to the BULLPEN: Sample-Efficient LLM Profiling via Multi-Channel Model Embeddings.¹

Their BULLPEN framework converts a model’s responses to a shared collection of questions into a numerical profile, or model embedding. A prediction model can then use that profile to estimate characteristics that would otherwise require separate evaluations.

The principal finding is that actual answer content contains information that correctness-centered profiles miss. But the advantage depends on what the organization wants to predict.

Answer text changes what a model profile can measure

BULLPEN constructs representations from three information channels:

1. Correctness bits: whether the model answers each question correctly.
2. Question text: numerical representations of questions, weighted by how much the model’s correctness differs from the cross-model average.
3. Answer text: numerical representations of the model’s generated responses.

The answer-text channel uses a pretrained sentence encoder, BGE-base-en-v1.5, to represent responses in 768 dimensions. Averaging those representations produces a reusable model-level profile.

BULLPEN also constructs a 32-dimensional fused representation using multi-block partial least squares and singular value decomposition. This method identifies directions that covary across channels, then combines standardized projected scores.

For restricted evaluation budgets, the framework selects questions using Fisher information from a fitted two-parameter item response theory model. Questions receive higher priority when their estimated difficulty and discrimination make them informative about differences among models.

The mechanism is straightforward: correctness preserves performance patterns, while response text can preserve tendencies in phrasing, verbosity, hedging, and refusal. Combining channels offers a way to retain both kinds of information, although the resulting representation need not be optimal for every prediction target.

Behavioral prediction improves when answers are retained

The researchers evaluate 260 open-weight checkpoints on 46 downstream tasks. Twenty are newly constructed behavior, style, and provenance tasks; the other 26 assess established competence and preference outcomes.

The new tasks include answer length, hedging, answer divergence, style effects, refusal frequency, model authorship, architecture, family, checkpoint specifications, and adoption indicators.

In the strict evaluation, models are divided into five organization-grouped folds, repeated across three experimental seeds. Encoders and prediction heads are fitted without the held-out fold. Evaluation questions are also separated from the questions used to construct embeddings.

The main results reveal a substantial difference between prediction objectives.

Sources: Section 5.1, Table 3; Section 5.2. These are normalized predictive skill values above width-matched random-embedding floors, not percentages of correct predictions.

The answer-only representation exceeds a correctness-only PCA baseline by 0.21 skill units on the 20 new tasks, with a paired interval of [0.12, 0.30]. It outperforms correctness-based encoders on 13 of those tasks, although intervals exclude zero on only six.

The strongest advantages concern answer statistics, refusal rate, and authorship. The pattern suggests that answer content is particularly informative for properties directly reflected in generated language.

However, the model’s success at predicting a refusal statistic does not mean it has been subjected to a comprehensive safety assessment. Likewise, predicting authorship or family is not definitive verification of training provenance.

The answer-swap control tests where the signal comes from

An alternative explanation is that a generic text encoder provides useful semantic features regardless of which model generated the responses.

The authors test this possibility by substituting another model’s answers when constructing the target model’s representation.

For BULLPEN-ans-full, this substitution reduces aggregate skill by approximately 0.46, with an interval of [0.40, 0.52], bringing performance close to the random-embedding floor.

The fused representation loses approximately 0.07 [0.03, 0.11], because correctness and question-text information remain available.

This is a robustness control, not a second independent deployment experiment. Its contribution is to show that the behavioral advantage depends materially on the model’s own outputs rather than merely on feeding answer-shaped text into an embedding system.

More information channels do not guarantee a better prediction

The competence results complicate the case for replacing correctness-based profiles.

Across 26 established competence and preference tasks, JE-IRT achieves 0.65 ± 0.08, compared with 0.55 ± 0.07 for BULLPEN’s fused representation. On all 46 tasks, the fused representation reaches 0.50 against JE-IRT’s 0.52; their paired difference is not statistically distinguishable from zero.

The fused embedding is competitive across a broader range of properties, but it is not uniformly superior.

Appendix G examines this more directly. Across 31 tested tasks, none satisfies the corrected criterion requiring the joint representation to outperform both the stronger single-channel alternative and its own shuffled-answer control.

This is a negative test of demonstrated fusion superiority, not proof that joint representations can never help. It means the study does not establish a reliable task-level gain from combining channels once the appropriate comparisons and multiple-testing correction are applied.

The practical inference is to choose profiling signals according to the intended prediction. A team concerned with benchmark competence has reason to retain correctness-based methods. A team concerned with response style or refusal patterns has reason to preserve answer text. A fused representation can summarize both, but its extra complexity requires task-specific justification.

Four hundred questions recover much of the measured skill

A richer behavioral profile would have limited practical appeal if constructing it required an exhaustive evaluation for every candidate.

BULLPEN therefore examines how predictive skill changes when its probe is restricted to selected questions.

The study’s original probe contains 31,286 questions. Of the 260 models, 104 have complete responses and 156 have responses to a smaller subset. The strict 260-model comparison uses a common pool of 900 questions, whereas the separate budget experiment uses the 104 fully evaluated models and compares restricted budgets against a 24,118-question pool.

At 400 selected questions, Appendix B, Table A4 reports:

The selected questions account for approximately 1.3% of full-pool inference compute under the authors’ accounting.

The 102% result reflects seed-level variation, not a demonstrated benefit from discarding more questions. More importantly, this efficiency analysis uses a faster protocol that does not hold out encoder-training models as strictly as the primary comparative evaluation.

The result supports economical profile construction within the tested setting. It does not establish that unseen production models will preserve identical predictive performance with 400 questions, or that total system cost falls proportionally once answer embedding, storage, and downstream fitting are included.

Model screening should separate competence from behavioral fit

For organizations comparing many open-weight checkpoints, BULLPEN suggests a practical screening architecture.

First, construct competence-oriented profiles to identify candidates likely to meet the application’s task requirements. Preserve actual correctness evaluations for important domain-specific requirements rather than substituting predicted scores for every direct measurement.

Next, construct answer-text profiles for shortlisted candidates. Use them to estimate response characteristics such as verbosity, hedging, style, and refusal frequency, prioritizing those that affect the intended user experience.

Finally, direct more expensive testing toward unresolved differences. A customer-support application might examine refusals on legitimate support requests; a coding application might investigate performance on representative development tasks. Predicted profiles would determine where to inspect more closely, not automatically authorize deployment.

This workflow is a Cognaptus operational inference, not a deployment outcome tested by the paper. Its value lies in making the screening decision more informative while potentially reducing redundant inference expenditure.

The remaining boundary is deployment validity

The comparative findings are reasonably supported within their experimental setting, but several boundaries remain.

All evaluated checkpoints are open-weight models answering single-turn questions. API-only models, multi-turn conversations, tool use, and autonomous agents are not tested. Only one pretrained answer-text encoder is examined, and generation configurations are not systematically varied.

Some downstream targets have sparse public labels, and some behavior measures depend on heuristic pattern matching. Organization-grouped folds also cannot guarantee that related checkpoints sharing an underlying base model are separated.

These constraints affect the interpretation of predicted identity, refusal behavior, and deployment reliability. A profile can reveal statistical tendencies in benchmark responses without establishing comprehensive safety, future behavior under distribution shift, or the success of production routing decisions.

The paper also does not evaluate whether those model profiles remain stable as model populations, checkpoints, or operating conditions change.

The selection criterion determines the right profile

BULLPEN’s strongest contribution is not that three information channels should always be combined. It demonstrates that the content of a model’s answers can be a valuable profiling signal in its own right, particularly when selection depends on behavior and provenance-related characteristics rather than competence alone.

The sample-efficiency results make this approach operationally plausible for large open-weight model collections. The unresolved fusion advantage argues against assuming that greater representational breadth automatically produces better decisions.

A model-selection process should therefore begin by defining what must be predicted, retain the information channels relevant to that target, and reserve direct evaluation for properties whose errors would materially affect deployment.

Accuracy is part of a model’s profile. Its answers contain more of the story.

Cognaptus: Automate the Present, Incubate the Future.