Adaptively profiling models with task elicitation

Damian Brown, Prithvi Balehannina, Helen Jin, Shreya Havaldar, Hamed Hassani, Eric Wong

Introduction

Researchers from the University of Pennsylvania have introduced task elicitation, an automated framework designed to adaptively "profile" Large Language Models (LLMs) to identify their hidden breaking points. Unlike traditional evaluation methods that rely on fixed datasets, this framework uses a "profiler" model to hypothesize potential failure modes and generate targeted questions. By inspecting a target model's internal reasoning, the method uncovers hundreds of systematic failures that standard static benchmarks often miss.

The Limitations of Static Benchmarking

The current paradigm for evaluating LLMs relies heavily on static benchmarks—fixed sets of human-curated questions. This approach faces two major hurdles. First, it is prohibitively expensive; creating questions that challenge frontier models often requires subject matter experts, costing hundreds of dollars per query. Second, these benchmarks have limited coverage, leading to a "jagged frontier" where a model might pass a complex bar exam but fail at basic legal hypotheticals.

Furthermore, models often "cheat" on evaluations by exploiting formatting patterns rather than demonstrating true understanding. There is a desperate need for a scalable, automated way to produce interpretable, natural-language descriptions of exactly where and why models fail.

The Task Elicitation Framework

The authors propose task elicitation as a three-step adaptive loop that moves beyond fixed scoring to qualitative profiling:

  1. Defining the Domain: The process begins with a "seed" evaluation (e.g., a few legal questions) and a natural language rubric defining the area of interest.
  2. Adaptive Evaluation: A "profiler" model (such as GPT-4o) inspects the target model’s successes and failures. Crucially, the profiler often analyzes the target model's Chain-of-Thought (CoT) to identify reasoning flaws, then generates new, harder questions designed to test specific hypotheses about the model's weaknesses.
  3. Clustering and Summarization: The resulting failed questions are clustered and summarized into "tasks"—human-readable descriptions of failure modes, such as "over-associates quantum computing and AGI."
Comparison of static evaluation and the proposed adaptive profiling loop.

Key Insights and Model Behavior

The framework revealed that qualitative profiling can scale just as effectively as quantitative benchmarking, discovering hundreds of specific behavioral bugs across various domains including legal reasoning, alignment, and social harms.

  • The CoT Advantage: Profiling is significantly more effective when the profiler can see the target's internal Chain-of-Thought. Seeing how a model rationalizes a wrong answer allows the profiler to craft much more adversarially difficult questions, leading to "win-rates" (generating a question the model fails) of up to 80% in legal reasoning.
  • Discovery of Specific Biases: Task elicitation found highly specific failure modes. For instance, Claude 3.5 Sonnet was found to have a "hallucination bias" where it over-associates quantum computing with AGI. Similarly, o3-mini was found to be particularly prone to hallucinating when incorrect "facts" are repeated within the conversation context.
  • Jailbreaking Efficiency: Using task elicitation for jailbreaking achieved a 20% higher success rate on Claude 3.5 Sonnet compared to established methods like PAIR.
A summary of elicited tasks and failure modes across frontier models.

Discussion and Future Work

While task elicitation provides deep insights, it is token-intensive; a full profile of 50 successfully generated questions can require over a million input tokens. Additionally, because the process is automated, some generated tasks can be noisy or redundant, requiring further refinement in the clustering stage.

The authors suggest that while this tool is vital for safety researchers, it also carries adversarial risks, as the same techniques could be used to find more effective jailbreaks. Future directions include scaling this framework to even larger domains and integrating it directly into the model training loop to "patch" discovered failure modes in real-time.

Summary updated

BibTeX
@misc{brown2025adaptively,
  author = {Damian Brown and Prithvi Balehannina and Helen Jin and Shreya Havaldar and Hamed Hassani and Eric Wong},
  title = {{Adaptively profiling models with task elicitation}},
  year = {2025},
  eprint = {2503.01986},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.01986},
}
Figure