Retrieval vs comprehension in long-context LLMs: a transfer study using MRCR and LongBench v2
Azzi, Leonardo (A.A. 2024/2025) Retrieval vs comprehension in long-context LLMs: a transfer study using MRCR and LongBench v2. Tesi di Laurea in Artificial intelligence and machine learning, Luiss Guido Carli, relatore Giuseppe Francesco Italiano, pp. 37. [Bachelor's Degree Thesis]
|
PDF (Full text)
Download (742kB) | Preview |
Abstract/Index
Background. Long-context models and why evaluation matters. Retrieval versus comprehension. The MRCR training signal. LongBench v2 as the transfer target. Why transfer is non-trivial in practice. Design principles used in this thesis. What this background implies for the rest of the thesis. Why MRCR and why LongBench v2 for this study. Inference‑time reasoning and long tasks. Methods. Model and software stack. Training data: MRCR. Training objective. Evaluation benchmark and protocol. Metrics. Statistical analysis. Reproducibility and provenance. Experiments and results. Reporting conventions. Overall accuracy and deltas. Bucketed results by length and difficulty. Transitions from baseline to MRCR SFT. Error patterns and formatting diagnostics. Domain‑level accuracy.
References
Bibliografia: pp. 28-30.
| Thesis Type: | Bachelor's Degree Thesis |
|---|---|
| Institution: | Luiss Guido Carli |
| Degree Program: | Bachelor's Degree Programs > Bachelor's Degree Program in Management and Computer Science, English language (L-18) |
| Chair: | Artificial intelligence and machine learning |
| Thesis Supervisor: | Italiano, Giuseppe Francesco |
| Academic Year: | 2024/2025 |
| Session: | Autumn |
| Deposited by: | Alessandro Perfetti |
| Date Deposited: | 29 Apr 2026 12:40 |
| Last Modified: | 29 Apr 2026 12:40 |
| URI: | https://tesi.luiss.it/id/eprint/45584 |
Downloads
Downloads per month over past year
Repository Staff Only
 |
View Item |
