国家预印本平台

Microfluidic devices have emerged as powerful tools in various laboratory applications, but the complexity of their design limits accessibility for many practitioners. While progress has been made in microfluidic design automation (MFDA), a practical and intuitive solution is still needed to connect microfluidic practitioners with MFDA techniques. This work introduces the first practical application of large language models (LLMs) in this context, providing a preliminary demonstration. Building on prior research in hardware description language (HDL) code generation with LLMs, we propose an initial methodology to convert natural language microfluidic device specifications into system-level structural Verilog netlists. We demonstrate the feasibility of our approach by generating structural netlists for practical benchmarks representative of typical microfluidic designs with correct functional flow and an average syntactical accuracy of 88%.

Large language models (LLMs) acquire a large amount of knowledge through pre-training on vast and diverse corpora. While this endows LLMs with strong capabilities in generation and reasoning, it amplifies risks associated with sensitive, copyrighted, or harmful content in training data. LLM unlearning, which aims to remove specific knowledge encoded within models, is a promising technique to reduce these risks. However, existing LLM unlearning methods often force LLMs to generate random or incoherent answers due to their inability to alter the encoded knowledge precisely. To achieve effective unlearning at the knowledge level of LLMs, we propose Knowledge Unlearning by Deviating representAtion (KUDA). We first utilize causal tracing to locate specific layers for target knowledge storage. We then design a new unlearning objective that induces the model's representations to deviate from its original position in the phase of knowledge removal, thus disrupting the ability to associate with the target knowledge. To resolve the optimization conflicts between forgetting and retention, we employ a relaxation null-space projection mechanism to mitigate the disruption to the representation space of retaining knowledge. Extensive experiments on representative benchmarks, WMDP and MUSE, demonstrate that KUDA outperforms most existing baselines by effectively balancing knowledge removal and model utility retention.

We study the genealogy of a sample of $k>1$ particles, taken under various sampling schemes, from the population alive at fixed times, in a continuous-time multitype Bienaymé-Galton-Watson (MBGW) tree with finite second moments. For critical MBGW trees under uniform sampling without replacement, we show that the sample genealogy converges in the large time limit to a universal limiting structure with the same tree topology as Kingman's coalescent but enriched by a nontrivial process coding the types along each lineage. Remarkably, the limiting genealogy is robust to the sampling scheme: it remains unchanged when fixing type configurations (that is, if we sample uniformly at random without replacement given a fixed vector of types to sample), or sampling with type-dependent weights (that is, multinomial sampling over types). Tracking types along the genealogy reveals that, in the limit, types evolve independently of the tree structure but retain a distribution determined by the offspring law, and the Perron-Frobenius eigenvalues. For uniform sampling without replacement, we use $k$ distinguished \emph{spine} particles and a suitable change of measure under which, when sampling at fixed times: (a) the spines form a uniform sample without replacement that depend on the types, and (b) there is $k$-size biasing and discounting according to the population size. This work substantially extends the spine techniques developed by Harris, Johnston, and Roberts 2020 and Harris, Johnston, and Pardo 2024, for the single-type case. In particular, we provide a detailed analysis of how type information affects functionals of the spines. We show that the multitype case introduces complex interactions between types, resulting in a richer dependency structure where functionals must capture type-specific behaviours and inter-type correlations.

Let $f$ be a Rademacher or Steinhaus random multiplicative function. For various arithmetically interesting subsets $\mathcal A\subseteq [1, N]\cap\mathbb N$ such that the distribution of $\sum_{n\in \mathcal A} f(n)$ is approximately Gaussian, we develop a general framework to understand the large fluctuations of the sum. This extends the general central limit theorem framework of Soundararajan and Xu. In the case when $\mathcal A = (N-H, N]$ is a short interval with admissible $H=H(N)$, we show that almost surely \begin{equation*} \limsup_{N\to\infty} \frac{\big\lvert\sum_{N-H<n\leq N} f(n)\big\rvert}{\sqrt{H\log \frac{N}H{}}}>0. \end{equation*} When $\mathcal A$ is the set of values of an admissible polynomial $P\in\mathbb Z[x]$, we extend work of Klurman, Shkredov, and Xu, as well as Chinis and the author, showing that almost surely \begin{equation*} \limsup_{N\to\infty} \frac{\big\lvert\sum_{n\leq N} f(P(n))\big\rvert}{\sqrt{N \log\log N}}>0, \end{equation*} even when $P$ is a product of linear factors over $\mathbb Q$. In this case, we also establish the corresponding almost sure upper bound, matching the law of iterated logarithm. An important ingredient in our work is bounding the Kantorovich--Wasserstein distance by means of a quantitative martingale central limit theorem.

The goal of the BabyLM is to stimulate new research connections between cognitive modeling and language model pretraining. We invite contributions in this vein to the BabyLM Workshop, which will also include the 4th iteration of the BabyLM Challenge. As in previous years, the challenge features two ``standard'' tracks (Strict and Strict-Small), in which participants must train language models on under 100M or 10M words of data, respectively. This year, we move beyond our previous English-only pretraining datasets with a new Multilingual track, focusing on English, Dutch, and Chinese. For the workshop, we call for papers related to the overall theme of BabyLM, which includes training efficiency, small-scale training datasets, cognitive modeling, model evaluation, and architecture innovation.