Abstract<br> Time-multiplexed networks of degenerate optical parametric oscillators have demonstrated remarkable success in simulating coupled Ising spins, thus providing a promising route to solving complex combinatorial optimization problems. In these systems, referred to as coherent Ising machines, spins are encoded in the oscillator phases, and measured at the system output using phase-sensitive techniques, making intricate phase stabilization necessary. Here, we introduce an optical Ising machine based on spontaneous polarization symmetry breaking in a coherently driven fiber Kerr nonlinear resonator. In our architecture, the spins are encoded in the polarization state, allowing robust, all-intensity readout with off-the-shelf telecom components. By operating in a newly-discovered regime where nonlinearity and topology lock the system’s symmetry, we eliminate drift and bias, enabling uninterrupted Ising trials at optical speeds for over an hour, without manual intervention. This all-fiber platform not only simplifies the hardware but also opens a path to more stable, high-throughput coherent optical optimization devices for applications from finance to drug design and beyond. <<<

AbstractBreast cancer is a highly heterogeneous disease, and there are many forms of categorization for breast cancer based on gene expression profiles. Gene expression profiles are variables and may show differences if measured at different time points or under different conditions. In contrast, biological networks are relatively stable over time and under different conditions. In this study, we used a gene interaction network from a new point of view to explore the subtypes of breast cancer based on individual-specific edge perturbations measured by relative gene expression value. Our study reveals that there are four breast cancer subtypes based on gene interaction perturbations at the individual level. The new network-based subtypes of breast cancer show strong heterogeneity in prognosis, somatic mutations, phenotypic changes and enriched pathways. The network-based subtypes are closely related to the PAM50 subtypes and immunohistochemistry index. This work helps us to better understand the heterogeneity and mechanisms of breast cancer from a network perspective. <<<

When receiving a reward after a sequence of multiple events, how do we determine which event caused the reward? This problem, known as temporal credit assignment, can be difficult for humans to solve given the temporal uncertainty in the environment. Research to date has attempted to isolate dimensions of delay and reward during decision-making, but algorithmic solutions to temporal learning problems and the effect of uncertainty on learning remain underexplored. To further our understanding, we adapted a reward learning task that creates a temporal credit assignment problem by combining sequentially delayed rewards, intervening events, and varying uncertainty via the amount of information presented during feedback. Using computational modeling, two learning strategies were developed: an eligibility trace, whereby previously selected actions are updated as a function of the temporal sequence, and a tabular update, whereby only systematically related past actions (rather than unrelated intervening events) are updated. We hypothesized that reduced information uncertainty would correlate with increased use of the tabular strategy, given the model’s capacity to incorporate additional feedback information. Both models effectively learned the task, and predicted choices made by participants (N = 142) as well as specific behavioral signatures of credit assignment. Consistent with our hypothesis, the tabular model outperformed the eligibility model under low information uncertainty, as evidenced by more accurate predictions of participants’ behavior and an increase in tabular weight. These findings provide new insights into the mechanisms implemented by humans to solve temporal credit assignment and adapt their strategy in varying environments. <<<

Recent methodologies in LLM self-training mostly rely on LLM generating responses and filtering those with correct output answers as training data. This approach often yields a low-quality fine-tuning training set (e.g., incorrect plans or intermediate reasoning). In this paper, we develop a reinforced self-training approach, called ReST-MCTS*, based on integrating process reward guidance with tree search MCTS* for collecting higher-quality reasoning traces as well as per-step value to train policy and reward models. ReST-MCTS* circumvents the per-step manual annotation typically used to train process rewards by tree-search-based reinforcement learning: Given oracle final correct answers, ReST-MCTS* is able to infer the correct process rewards by estimating the probability this step can help lead to the correct answer. These inferred rewards serve dual purposes: they act as value targets for further refining the process reward model and also facilitate the selection of high-quality traces for policy model self-training. We first show that the tree-search policy in ReST-MCTS* achieves higher accuracy compared with prior LLM reasoning baselines such as Best-of-N and Tree-of-Thought, within the same search budget. We then show that by using traces searched by this tree-search policy as training data, we can continuously enhance the three language models for multiple iterations, and outperform other self-training algorithms such as ReST$^\text{EM}$ and Self-Rewarding LM. We release all code at https://github.com/THUDM/ReST-MCTS. <<<

Current approaches to reducing undesired capabilities in language models are largely post hoc, and can thus be easily bypassed by adversaries. A natural alternative is to shape capabilities during pretraining itself. On the proxy task of removing medical capabilities, we show that the simple intervention of filtering pretraining data is highly effective, robust, and inexpensive at scale. Inspired by work on data attribution, we show that filtering tokens is more effective than filtering documents, achieving the same hit to undesired capabilities at a lower cost to benign ones. Training models spanning two orders of magnitude, we then demonstrate that filtering gets more effective with scale: for our largest models, token filtering leads to a 7000x compute slowdown on the forget domain. We also show that models trained with token filtering can still be aligned on the forget domain. Along the way, we introduce a methodology for labeling tokens with sparse autoencoders and distilling cheap, high-quality classifiers. We also demonstrate that filtering can be robust to noisy labels with sufficient pretraining compute. <<<

AbstractDiminishing natural resources and increasing climatic volatility are impacting agri-food systems, prompting the need for sustainable and resilient alternatives. Python farming is well established in Asia but has received little attention from mainstream agricultural scientists. We measured growth rates in two species of large pythons (<i>Malayopython reticulatus</i> and <i>Python bivittatus</i>) in farms in Thailand and Vietnam and conducted feeding experiments to examine production efficiencies. Pythons grew rapidly over a 12-month period, and females grew faster than males. Food intake and growth rates early in life were strong predictors of total lifetime growth, with daily mass increments ranging from 0.24 to 19.7 g/day for <i>M. reticulatus</i> and 0.24 to 42.6 g/day for <i>P. bivittatus</i>, depending on food intake. Pythons that fasted for up to 4.2 months lost an average of 0.004% of their body mass per day, and resumed rapid growth as soon as feeding recommenced. Mean food conversion rate for dressed carcasses was 4.1%, with useable products (dressed carcass, skin, fat, gall bladder) comprising 82% of the mass of live animals. In terms of food and protein conversion ratios, pythons outperform all mainstream agricultural species studied to date. The ability of fasting pythons to regulate metabolic processes and maintain body condition enhances food security in volatile environments, suggesting that python farming may offer a flexible and efficient response to global food insecurity. <<<

Deep learning models that predict functional genomic measurements from DNA sequences are powerful tools for deciphering the genetic regulatory code. Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance. We present AlphaGenome, a unified DNA sequence model, which takes as input 1 Mb of DNA sequence and predicts thousands of functional genomic tracks up to single-base-pair resolution across diverse modalities. The modalities include gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage and splice junction coordinates and strength. Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction. The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene. To facilitate broader use, we provide tools for making genome track and variant effect predictions from sequence. <<<

Abstract<br> Inexpensive and accurate genotyping methods are essential to modern genomics and health risk prediction. Here we introduce QUILT2, a scalable and read-aware imputation method that can efficiently use biobank scale haplotype reference panels. This allows for fast and accurate imputation using short reads, as well as long reads (e.g. Oxford Nanopore Technologies (ONT) 1X, r2 = 0.937 at common SNPs), linked-reads and ancient DNA. In addition, QUILT2 contains a methodological innovation that is designed to enable imputation of the maternal and fetal genome using cell free non-invasive prenatal testing (NIPT) data. Using a UK Biobank reference panel and simulated NIPT data, we see accurate imputation of the mother (0.25X, r2 = 0.966, common SNPs) and modest imputation of the fetus (0.25X, r2 = 0.465, fetal fraction of 10%) at low coverage, with fetal imputation accuracy rising with coverage (4.0X, fetal r2 = 0.894). We show using simulated data that this could enable both GWAS and PRS for the mother and fetus, which could create clinical opportunities, and if phenotypes can be collected alongside clinical NIPT, the potential for large GWAS. <<<

Abstract<br> Wheat (Triticum aestivum L.) spike development is tightly regulated by genetic and metabolic programs that drive organ growth and morphological changes. However, the dynamic interplay between metabolic shifts, gene expression patterns, and their regulatory roles during spike development, remains poorly characterized. To address this knowledge gap, we performed integrated metabolomic and transcriptomic profiling across 12 stages of wheat spike organ development. Our analysis detected 1,105 metabolites in 233 spike, spikelet, and floret samples, uncovering an uneven distribution of phytohormone-related metabolites. The exogenous phytohormone treatments validated the regulatory roles of phytohormones in spike morphogenesis. High-resolution spatiotemporal data from carpel organs enabled the reconstruction of a regulatory network, identifying key genes (including 12-oxo-phytodienoic acid reductase3 (TaOPR3), Grain Length1 (GL1), and Grain Length2 (GL2)) as critical determinants of grain size. Genomic analyses revealed geographical differentiation in gene haplotypes and their selective retention during breeding, with superior alleles associated with increased grain size. This comprehensive dataset provides a valuable resource for understanding the molecular basis of wheat grain yield and offers potential targets for crop improvement. <<<

We explore how generating a chain of thought -- a series of intermediate reasoning steps -- significantly improves the ability of large language models to perform complex reasoning. In particular, we show how such reasoning abilities emerge naturally in sufficiently large language models via a simple method called chain of thought prompting, where a few chain of thought demonstrations are provided as exemplars in prompting. Experiments on three large language models show that chain of thought prompting improves performance on a range of arithmetic, commonsense, and symbolic reasoning tasks. The empirical gains can be striking. For instance, prompting a 540B-parameter language model with just eight chain of thought exemplars achieves state of the art accuracy on the GSM8K benchmark of math word problems, surpassing even finetuned GPT-3 with a verifier. <<<

Recent breakthroughs in protein structure prediction have opened new avenues for genome-wide drug discovery, yet existing virtual screening methods remain computationally prohibitive. We present DrugCLIP, a contrastive learning framework that achieves ultrafast and accurate virtual screening, up to 10 million times faster than docking, while consistently outperforming various baselines on in silico benchmarks. In wet-lab validations, DrugCLIP achieved a 15% hit rate for norepinephrine transporter, and structures of two identified inhibitors were determined in complex with the target protein. For thyroid hormone receptor interactor 12, a target that lacks holo structures and small-molecule binders, DrugCLIP achieved a 17.5% hit rate using only AlphaFold2-predicted structures. Finally, we released GenomeScreenDB, an open-access database providing precomputed results for ~10,000 human proteins screened against 500 million compounds, pioneering a drug discovery paradigm in the post-AlphaFold era. <<<