Recent Visual-Language Models (VLMs) have enhanced the capabilities of pre-trained LLMs by adding vision tokens alongside text, with approaches like LLaVA showing impressive results. However, the computational burden of processing up to 576 or 729 visual tokens makes edge deployment challenging. While various token pruning techniques require retraining, some are training-free and thus can easily adapt to architecture changes. We introduce ClustRS, a two-part, training-free algorithm for robust token pruning. Its first component is an attention-weighted, clustering algorithm that selects representative tokens from each semantic cluster. The second component, Residual Shrinkage, is a one-pass denoising step on the selected tokens. These training-free lightweight steps make LLaVA ready for real-world data, improving robustness to a wide range of image-noise types and intensities. Experimental results on the ScienceQA-IMG and MM-VET benchmarks show our method outperforms attention- and diversity-based methods by up to 20% under extreme noise and token conditions (reducing tokens by 97%, down to 16 tokens) on LLaVA 1.5 7b and achieves better results on LLaVA-OneVision, where we match baseline performance with fewer than one-third of their tokens under mild noise conditions. Our study demonstrates a simple yet powerful alternative to both score-only and diversity-only pruning rules, paving the way for compute-efficient and noise-resilient VLM deployment.
Perihilar cholangiocarcinoma (CCA) is an aggressive malignancy that presents significant challenges in terms of diagnosis and treatment owing to its anatomical location. Liver resection and transplantation are the primary curative options; however, their long-term outcomes remain controversial. This study aimed to compare the long-term outcomes of protocol-based neoadjuvant selection and liver transplantation with liver resection for the management of perihilar cholangiocarcinoma. A systematic review and meta-analysis were conducted according to the PRISMA guidelines. Electronic databases, including the Cochrane Library, PubMed/MEDLINE, and Embase, were searched for relevant clinical trials comparing liver transplantation with liver resection. Six studies comprising 768 patients (265 in the liver transplantation group and 503 in the liver resection group) were included. Liver transplantation was associated with significantly higher 5-year overall survival and 5-year disease-free survival. They also showed a higher 3-year disease-free survival rate. Patients in the liver transplantation group had a higher R0 resection rate. There were no significant differences in the 3- or 30-day postoperative mortality between the two groups. Liver transplantation was associated with improved long-term survival in highly selected patients; however, these findings are limited by significant selection bias and the retrospective study design. The apparent long-term benefit observed with LT must be interpreted cautiously because transplant series frequently report post-transplant rather than intention-to-treat outcomes.
PURPOSE:Cancer cachexia affects up to 80% of patients with advanced cancer. While exercise has shown benefits for early-stage cancer patients, its role in managing cachexia remains unclear. This study examined whether severe weight loss impedes the benefits of exercise in cancer patients. METHODS:This prospective study analyzed the feasibility and efficacy of a 12-week supervised exercise program in cancer patients with BMI-adjusted weight loss grades 3-4. Feasibility was the primary outcome and was evaluated through attrition, attendance, and safety, while secondary outcomes included changes in nutritional status, physical performance, exercise levels, and fatigue. Outcomes were compared with those of matched cancer patients without advanced BMI-adjusted weight loss who followed the same exercise program. RESULTS:Twenty-three patients with BMI-adjusted weight loss grades 3-4 were enrolled and 14 completed the program, attending an average of 17.1 ± 4.3/24 sessions. No adverse events occurred. Intervention improved endurance (+7.06% of theoretical 6MWT distance, p < 0.001). Positive trends emerged for lower limb strength (+2.69 repetitions at 30-s sit-to-stand test, p = 0.005), lean mass (+1.78 kg, p = 0.003), and dietary intake (+ 2.5/10 SEFI, p = 0.003). Functional benefits were also observed in patients without weight loss (n = 14, endurance = +11.33%, lower limb strength = +4, p < 0.001). CONCLUSION:The exercise intervention demonstrated promising efficacy in managing cancer weight loss, providing benefits like those observed in non-cachectic patients, with added gains in nutritional status. While no exercise-related adverse events occurred, feasibility criteria were only partially met due to a higher-than-expected attrition rate. Further research is needed to optimize recruitment strategies, energetic balance, and tailor interventions for this population. TRIAL REGISTRATION:The study is registered on ClinicalTrial.gov (Trial registration: NCT06651125).
Covering problems in coding theory are closely related to finite geometry through the interpretation of the columns of parity-check matrices as point sets in finite vector spaces. Motivated by the recent notion of generalized covering radii of linear codes introduced by Elimelech, Firer and Schwartz, we develop a geometric framework for these parameters. We introduce $(ρ,t)$-saturating sets and show that they are precisely the finite-geometric counterparts of linear codes whose $t$-th generalized covering radius is at most $ρ$. We study the structure of these sets and show that the extremal case $ρ=t$ coincides with the notion of $t$-strong blocking sets. Thus, $(ρ,t)$-saturating sets interpolate between classical saturating sets and strong blocking sets. We provide several equivalent formulations, including affine and dual Grassmannian criteria, derive lower bounds on their size, and give constructions from strong blocking sets, graphs and projective configurations.
How do start-ups create value through the language of their business pitches? In this article, we investigate that question by identifying the logics of justification they use, traditionally conceptualized as orders of worth. In this study of short written pitches in a 6-month Chilean accelerator program, we describe how we detected logics of justification through pitch language, and we identify (a) co-occurrence patterns among logics of justification, (b) associations between logics of justification and industry sectors, and (c) associations between logics and a firm’s customer segment (B2B, B2C). This study provides unique insights into how start-ups sometimes justify innovations by using specific patterns of language depending on a venture’s features.
