By Noam Razin, Zixuan Wang, Hubert Strauss, Stanley Wei, Jason D. Lee, Sanjeev Arora

The success of Reinforcement Learning from Human Feedback (RLHF) critically depends on the quality of the reward model. However, while this quality is primarily evaluated through accuracy, it remains unclear whether accuracy fully captures what makes a reward model an effective teacher. We address this question from an optimization perspective. First, we prove that regardless of how accurate a reward model is, if it induces low reward variance, then the RLHF objective suffers from a flat landscape. Consequently, even a perfectly accurate reward model can lead to extremely slow optimization, underperforming less accurate models that induce higher reward variance. We additionally show that a reward model that works well for one language model can induce low reward variance, and thus a flat objective landscape, for another. These results establish a fundamental limitation of evaluating reward models solely based on accuracy or independently of the language model they guide. Experiments using models of up to 8B parameters corroborate our theory, demonstrating the interplay between reward variance, accuracy, and reward maximization rate. Overall, our findings highlight that beyond accuracy, a reward model needs to induce sufficient variance for efficient~optimization.

Read the paper: https://arxiv.org/abs/2503.15477

By Luis Linares-Otoya, Jaden D. Shirkey, Bhuwan Khatri Chhetri, Amira Mira, Abhishek Biswas, Samuel L. Neff, Maria V. Linares-Otoya, Ye Chen, Julio V. Campos-Florian, Mayar L. Ganoza-Yupanqui, Philip D. Jeffrey, Frederick M. Hughson & Mohamed S. Donia

Considerable advances have been made in characterizing bioactive molecules secreted by bacteria, yet the regulatory elements controlling their production remain largely understudied. Here we identify and characterize the N-acyl-cyclolysine (ACL) system—a cell-density-dependent chemical signalling system specific to and widespread in the phylum Bacteroidota (formerly Bacteroidetes)—and show that it regulates the expression of co-localized operons encoding diverse secreted molecules. Using genetic and biochemical analyses, combined with structural studies of a key biosynthetic enzyme, AclA, we elucidate the molecular structure of various ACLs and their complete biosynthetic pathway involving l-lysine acylation and ATP-dependent cyclization. Furthermore, we find that secreted ACLs are sensed by a dedicated transcription factor, AclR, resulting in the expression of associated operons and the autoinduction of ACL biosynthesis. Moreover, we show that different Bacteroidota strains produce structurally diverse ACLs and encode transcription factors with varying ligand specificities. Finally, we find that the acl circuit is widely distributed and transcribed in human gut and oral microbiome samples, with clear evidence for an active role in regulating associated operons under host colonization conditions. Understanding the function of the ACL system in different contexts has the potential to reveal details about the biology, ecology and chemistry of the Bacteroidota and how members of this phylum interact with their environments and hosts.

Read the paper: https://www.nature.com/articles/s41586-025-09418-9

By Jiayu Zhang, Elizabeth C. Chavez, Melina Winkler, Jianche Liu, Sebastian Carver, Aaron E. Lin, Abhishek Biswas, Tomokazu Tamura, Anna Tseng, Danyang Wang, Aaron Benhamou, Aoife K. O’ Connell, Mao Matsuo, Jack E. Norton, Devin Kenney, Britt Adamson, Ralph E. Kleiner, Benjamin Burwitz, Nicholas A. Crossland, Florian Douam & Alexander Ploss

The live-attenuated yellow fever 17D vaccine strain differs genetically only minimally from its virulent parent. However, it remains unclear which sequence differences lead to virulence or attenuation. Here we demonstrate, using SHAPE-MaP, that these mutations do not induce global RNA structure changes and show that protein sequence mutations are mostly responsible for the phenotypic differences between 17D and virulent YFV. Using a highly modular, combinatorial genetic approach, we identified key mutations in the envelope (E) and non-structural 2A (NS2A) proteins that increase 17D’s ability to spread and enhance host antiviral responses. Introducing these mutations into infectious clones of virulent YFV genomes results in viral attenuation in vitro and in two mouse models. Collectively, our results define the genetic basis for 17D attenuation and highlight a potentially general approach for creating live-attenuated vaccines by introducing mutations resulting in similar phenotypic changes in other pathogenic viruses.

Read the paper: https://www.nature.com/articles/s41564-025-02047-y

By Sofia A. Quinodoz, Lifei Jiang, Aya A. Abu-Alfa, Troy J. Comi, Hongbo Zhao, Qiwei Yu, Lennard W. Wiesner, Jordy F. Botello, Anita Donlic, Elizabeth Soehalim, Prashant Bhat, Christiane Zorbas, Ludivine Wacheul, Andrej Košmrlj, Denis L. J. Lafontaine, Sebastian Klinge & Clifford P. Brangwynne

Biomolecular condensates are key features of intracellular compartmentalization. As the most prominent nuclear condensate in eukaryotes, the nucleolus is a multiphase liquid-like structure in which ribosomal RNAs (rRNAs) are transcribed and processed, undergoing multiple maturation steps to form the small (SSU) and large (LSU) ribosomal subunits. However, how rRNA processing is coupled to the layered organization of the nucleolus is poorly understood owing to a lack of tools to precisely monitor and perturb nucleolar rRNA processing dynamics. Here we developed two complementary approaches to spatiotemporally map rRNA processing and engineer de novo nucleoli. Using sequencing in parallel with imaging, we found that rRNA processing steps are spatially segregated, with sequential maturation of rRNA required for its outward movement through nucleolar phases. By generating synthetic nucleoli in cells using an engineered rDNA plasmid system, we show that defects in SSU processing can alter the ordering of nucleolar phases, resulting in inside-out nucleoli and preventing rRNA outflux, while LSU precursors are necessary to build the outermost layer of the nucleolus. These findings demonstrate how rRNA is both a scaffold and substrate for the nucleolus, with rRNA acting as a programmable blueprint for the multiphase architecture that facilitates assembly of an essential molecular machine.

Read the paper: https://www.nature.com/articles/s41586-025-09207-4

By Naman Agarwal, Xinyi Chen, Evan Dogariu, Devan Shah, Hubert Strauss, Vlad Feinberg, Daniel Suo, Peter Bartlett, Elad Hazan

We address the challenge of efficient auto-regressive generation in sequence prediction models by introducing FutureFill, a general-purpose fast generation method for any sequence prediction algorithm based on convolutional operators. FutureFill reduces generation time from quadratic to quasilinear in the context length. Moreover, when generating from a prompt, it requires a prefill cache whose size grows only with the number of tokens to be generated, often much smaller than the caches required by standard convolutional or attention based models. We validate our theoretical claims with experiments on synthetic tasks and demonstrate substantial efficiency gains when generating from a deep convolutional sequence prediction model.

Read the paper: https://arxiv.org/abs/2410.03766

We’re excited to announce the publication of a new paper, Ten Simple Rules for Catalyzing Collaborations and Building Bridges between Research Software Engineers (RSEs) and Software Engineering Researchers (SERs), authored by Nasir Eisty, Jeffrey Carver, Johanna Cohoon, Ian Cosden, Carole Goble, and Samuel Grayson.

Published in IEEE Computing in Science & Engineering (CiSE), this work emerged from discussions at a Dagstuhl Seminar and addresses a critical but often overlooked opportunity in the research software ecosystem: fostering collaboration between RSEs and SERs.

While both communities share a passion for improving software in research, they often operate in distinct environments, with different vocabularies, incentives, and expectations. This paper offers ten actionable rules designed to bridge those gaps, encouraging meaningful, sustained partnerships that combine practical experience with theoretical insight.

By working together, RSEs and SERs can drive innovation in tools, practices, and infrastructure, ultimately advancing the quality and impact of scientific research.

Read the preprint: https://arxiv.org/abs/2506.03012

Published version: https://ieeexplore.ieee.org/document/11003859

By Ted Zadouri, Hubert Strauss, and Tri Dao

LLM decoding is bottlenecked for large batches and long contexts by loading the key-value (KV) cache from high-bandwidth memory, which inflates per-token latency, while the sequential nature of decoding limits parallelism. We analyze the interplay among arithmetic intensity, parallelization, and model quality and question whether current architectures fully exploit modern hardware. This work redesigns attention to perform more computation per byte loaded from memory to maximize hardware efficiency without trading off parallel scalability. We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfers without compromising model quality. We then introduce Grouped Latent Attention (GLA), a parallel-friendly latent attention paired with low-level optimizations for fast decoding while maintaining high model quality. Experiments show that GTA matches Grouped-Query Attention (GQA) quality while using roughly half the KV cache and that GLA matches Multi-head Latent Attention (MLA) and is easier to shard. Our optimized GLA kernel is up to 2x faster than FlashMLA, for example, in a speculative decoding setting when the query length exceeds one. Furthermore, by fetching a smaller KV cache per device, GLA reduces end-to-end latency and increases throughput in online serving benchmarks by up to 2x.

Read the paper: https://arxiv.org/abs/2505.21487

By Lindsay A. Becker, Sofia A. Quinodoz, Troy J. Comi, Ofer Kimchi, David A. Knowles, and Clifford P. Brangwynne

Subcellular RNA organization can affect critical cellular functions. However, our understanding of RNA microenvironments, particularly biomolecular condensates, remains limited, largely due to a lack of technologies to comprehensively interrogate mesoscale RNA organization. Here, we adapt Split-Pool Recognition of Interactions by Tag Extension to map micron-scale RNA-RNA spatial proximity genome-wide across cell regions (RNA-SPRITE). Deploying RNA-SPRITE, we find extensive, conserved organization of mature mRNAs, with increased colocalization between mRNAs that share RNA-binding protein (RBP) motifs or encode functionally related proteins. Both effects are especially strong in dendrites and axons, suggesting prevalent mRNA co-regulation. Moreover, mRNAs with less compact folding, lower translation efficiency, and specific RBP motifs are more likely to be in RNA-rich condensates. However, perturbations that broadly dissolve or enhance condensation reveal that RBP motif and encoded protein-mediated colocalizations largely remain intact, independent of condensation. These results demonstrate the power of RNA-SPRITE in revealing critical aspects of RNA’s functional organization.

In Brief Unbiased, genome-wide maps of RNA-RNA mesoscale spatial proximity uncover extensive subcellular organization and its governing principles.

Highlights

• RNA-SPRITE reveals micron-scale RNA colocalization genome-wide across cell regions
• mRNA colocalization specificity is driven by shared motifs and encoded protein function
• mRNAs with less compact folding, lower translation efficiency, and distinct protein-binding motifs are more likely to be in condensates
• Neurites have a particularly high degree of sequence and function-dependent mRNA organization

Read the paper: https://www.biorxiv.org/content/10.1101/2025.04.19.649570v1

By Mohamed Abdellatif, Joel U. Bretheim, and Marina Rustow

Different writing systems have been (historically and contemporarily) used to write out the same language. This is typically done by substituting letters (or symbols, in the case of non-alphanumeric systems). However, depending on the language and the involved writing systems, the process may not be purely deterministic. Quoting Becker and Becker [2000]

even such basic acts as transliteration involve interpretation– to the extent that there is
meaning in the medium itself
.
.
In transliteration itself there is exuberance (that is, meaning is added) and deficiency
(meaning is lost).

This gives significance to the problem of Machine Translation in the intersection of Digital Humanities and Natural Language Understanding. Transformer-based models achieved success modeling human languages. However, many of them have the limitation of handling an input of maximum length of 512 tokens. To reuse a pre-trained model with this limitation for downstream tasks (e.g., Machine Transliteration) on input of sequences longer than 512 tokens, we propose a method to segment the input into interleaving (not mutually exclusive) pieces, invoke the model in a piecewise manner and construct the result. To consolidate the result, we propose a method to detect and correct potential (duplication and elimination) errors that reduces Word Error Rate from 0.0985 to 0.0.

Read the paper: https://zenodo.org/records/14982300