CruxBench: A Benchmark of Information Discovery

Large language models (LLMs) are increasingly used to support decisions under incomplete information, where deciding what to look up and when to stop matters as much as the final answer. However, LLMs fall into an information trap: They pay for checks, i.e., costly pieces of evidence, that cannot change their decision or are not worth their cost, and they keep asking when the evidence is already sufficient. To this end, we propose Decision-Informed Value-of-information Acquisition (DIVA), a simple yet effective procedure for LLM information seeking based on the decision value of evidence. DIVA 1) lets the LLM specify a small decision model, i.e., how likely a bad outcome is and how each check would shift this probability, through short numeric questions and 2) values each candidate check by how much its imagined outcomes improve the decision net of its cost, stopping when no check is worth its cost. Its variant, DIVA-U, further accounts for the LLM's uncertain estimates of how strongly each check shifts this probability. DIVA is training-free and applies to any LLM. Across four domains and 20 LLMs, DIVA-U has lower regret, i.e., value lost relative to the best possible decision, than every LLM baseline on real-world data, mainly by stopping once evidence suffices, and DIVA cuts the regret of the best LLM baseline by 75% in synthetic scenarios where every check's reliability is given. Code: github.com/seunghan96/diva.

To use language models to help us make decisions we would like to understand the rationale behind the language model's outputs. In principle, the language model provides its rationale as part of its text output. However, the sheer volume of text makes this hard to use in practice. In this paper, we are interested in extracting the cruxes that lead a model to recommend one action over another. We take a crux to be a sub-question that a model's rollouts disagree on, where changing its value may have altered the conclusion. A crux is therefore a property of the distribution of a model's rollouts, not of any one rollout. We consider three strategies of increasing computational cost. First, we just ask a model to identify, a priori, what the cruxes of a decision will be. Second, we have the model actually go through the process of generating a response and then extract the cruxes that arose in its reasoning. Finally, we generate many responses and extract the cruxes by comparing the reasoning across the rollouts. We make this last procedure computationally tractable with cheap calls to a reranker, an estimated 86 times less compute than a judge. We then conduct an extensive empirical evaluation to assess whether the computationally inexpensive options are sufficient. They are not. Across domains and models, we find that a model's a priori assessment of the cruxes has little relation to the factors that are actually identified as important a posteriori. Using a single rollout helps, but is still inadequate for identifying the full set of cruxes. Accordingly, for high-stakes decisions (and to understand the space of language model decision making) the many rollout procedure is unavoidable.

Large language models (LLMs) are powerful black-box systems, making it difficult to discern whether their answers reflect stable internal beliefs or superficial pattern matching. We identify cross-contextual consistency as an underutilized behavioral property of LLMs: a credible answer should remain stable when the same task is placed under topic-aligned, content-neutral contextual variation. Building on this intuition, we operationalize Cross-Contextual Consistency (C3) by comparing model generations under original and perturbed prompts. Across 26 models and six benchmarks spanning reasoning, factuality, and code generation, we find that answers with smaller cross-contextual shifts are more likely to be correct or factual. We demonstrate that C3 provides a complementary axis of evaluation and can serve as a benchmark usefulness diagnostic, identifying which portions of a benchmark remain informative even when aggregated scores are widely considered "saturate".