Abstract:

[Purpose/Significance] In the digital intelligence era, the way college students seek academic information is shaped by two distinct pressures: the continuously growing volume of information, and the high cognitive investment required for querying, screening, verifying and integrating sources. The advent of large language models (LLMs), such as DeepSeek, in academic information retrieval has transformed the ways users acquire information, allowing users to bypass part of the conventional retrieval process and to access the model's integrated textual output directly. However, whether the efficiency gains commonly attributed to such tools are genuinely accompanied by a reduction in cognitive load cannot be judged from self-report alone, because subjective ratings are susceptible to social desirability and post hoc rationalization. This study therefore sets out to examine whether LLM-assisted retrieval reduces the cognitive load experienced by college students experience across academic search tasks of varying complexity, from both subjective and objective perspectives. [Method/Process] A two-factor mixed experimental design was adopted, crossing tool type (DeepSeek versus a conventional academic database, a between-subjects factor) and task complexity (low versus high, a within-subjects factor). Sixty college students, balanced across disciplines and gender, were recruited. Each completed both a low-complexity fact-finding task and a high-complexity review-type task, with task order counterbalanced to control for practice and fatigue effects. Subjective load was measured with the NASA-TLX scale, and objective load with a set of eye-tracking indicators, including baseline-corrected pupil diameter, total fixation time, fixations per unit time, regression rate, and dwell proportion in the target area of interest. Retrieval report quality was scored by two blind raters and further decomposed into model-output quality, student-processing quality and student-verification quality, so that any reduction in verification effort could be observed separately. The data were analyzed with a 2 × 2 mixed-design analysis of variance, with partial η² as the effect-size index and simple-effect analyses following significant interactions. [Results/Conclusions] Compared to the conventional database, retrieval with DeepSeek produced statistically significant reduction in subjective load (tool main effect F(1,58)=14.0, p<0.001, partial η²=0.19), a significantly smaller baseline-corrected pupil diameter, a shorter total fixation time, a lower regression rate, and a higher proportion of time spent in the target area of interest. Retrieval report quality did not decline. The moderating effect of task complexity was statistically significant for total fixation time and regression rate (interaction F=13.9 and 5.7, respectively), where DeepSeek's advantage was larger in high-complexity tasks, but was not statistically significant for subjective load or pupil diameter. Notably, although model-output quality was higher in the DeepSeek group, student verification quality was about three points lower than in the conventional group, indicating that the reduction in load was accompanied by a contraction of verification effort. LMMs help to relieve part of the extraneous cognitive load in academic information retrieval, mainly at the stages of information screening, repeated comparison and cross-text integration, while leaving largely unchanged the intrinsic load of comprehending content. The benefit is double-edged: cognitive offloading may weaken memory retention and critical verification. The paper discusses the measurement validity of eye-tracking indicators and offers suggestions for tool design and information literacy education, arguing that the evaluation of retrieval tools should extend beyond efficiency and load to encompass knowledge retention, source verification and critical processing.

Key words: large language model (LLM), DeepSeek, cognitive load, eye-tracking, information retrieval