Demystifying Scientific Problem-Solving in LLMs by Probing Knowledge and Reasoning

TL;DR

The paper introduces SciReas, a holistic benchmark suite for knowledge-intensive scientific reasoning across ten domains, and SciReas-Pro, a harder subset to better differentiate reasoning capabilities. It also presents KRUX, a probing framework that injects Knowledge Ingredients extracted from reasoning traces to disentangle the roles of knowledge and reasoning. Key findings show that retrieving task-relevant knowledge from model parameters is a bottleneck, external in-context knowledge benefits reasoning models, and verbalized reasoning improves knowledge surfacing. Together, these contributions offer a path toward more capable scientific reasoning systems through controlled data composition, external knowledge integration, and explicit reasoning traces.

Abstract

Scientific problem solving poses unique challenges for LLMs, requiring both deep domain knowledge and the ability to apply such knowledge through complex reasoning. While automated scientific reasoners hold great promise for assisting human scientists, there is currently no widely adopted holistic benchmark for evaluating scientific reasoning, and few approaches systematically disentangle the distinct roles of knowledge and reasoning in these tasks. To address these gaps, we introduce SciReas, a diverse suite of existing benchmarks for scientific reasoning tasks, and SciReas-Pro, a selective subset that requires more complex reasoning. Our holistic evaluation surfaces insights about scientific reasoning performance that remain hidden when relying on individual benchmarks alone. We then propose KRUX, a probing framework for studying the distinct roles of reasoning and knowledge in scientific tasks. Combining the two, we conduct an in-depth analysis that yields several key findings: (1) Retrieving task-relevant knowledge from model parameters is a critical bottleneck for LLMs in scientific reasoning; (2) Reasoning models consistently benefit from external knowledge added in-context on top of the reasoning enhancement; (3) Enhancing verbalized reasoning improves LLMs' ability to surface task-relevant knowledge.

Figures (18)

• Figure 1: KRUX pipeline. Starting upper left, we prompt an LLM (one of base, DeepSeek-R1, Base-Math, Base-STEM, and Base-BOTH) with a question from SciReas as the knowledge source, collect the output and reasoning traces, and feed the reasoning traces to DeepSeek-R1 as the extractor to generate knowledge ingredients (KIs). We then evaluate the tested model with KI-augmented questions, which allows us to study three key research questions (RQ1 §\ref{['sec:knowledge-recall']}, RQ2 §\ref{['sec:knowledge-usage']}, RQ3 §\ref{['sec:ver-recall']}) regarding LLMs' knowledge and reasoning capabilities in scientific problem-solving.
• Figure 2: Frontier reasoning models' performance evaluated on SciReas. The X-axis shows the cost per 1k instances in USD. Different reasoning settings on the same model can result in distinct costs and performance, but the margins vary depending on the models.
• Figure 3: An example pair with varying reasoning intensity, where the example on the left is sampled from SciReas-Pro and the right is a filtered out example (§\ref{['sec:scireasbench-pro']}). On the left, the progressive reasoning chain is highlighted. The example on the right emphasizes knowledge recall on each option with a simple elimination strategy.
• Figure 4: SciReas correlations breakdown. (a) Task-to-task Pearson correlations. SciReas incorporates tasks complementary to popular benchmarks. (b) and (c) show performance on SciReas vs. SciBench and MMLU-Pro*. Models may be tuned for certain tasks, outperforming higher-ranked models on individual benchmarks.
• Figure 5: Model performance on SciReas and SciReas-Pro with varying reasoning capabilities. SciReas-Pro amplifies gaps between low-reasoning and high-reasoning settings.