About

Jeff Gore is a Professor of Physics at the Massachusetts Institute of Technology who studies the emergent dynamics of complex communities, integrating experiment and theory to explore how interactions within populations or communities lead to rich emergent phenomena. His research focuses on laboratory microcosms to understand the loss of stability, chaotic fluctuations, alternative stable states, community assembly, and the spread of 'cheater' strategies. He has demonstrated key theoretical predictions about community stability and fluctuations through experimental work with microbial populations, including the effects of environmental modifications such as pH on microbial interactions. Gore's work on alternative stable states and tipping points has provided insights into ecological transitions, with experiments measuring early warning indicators like critical slowing down and spatial pattern emergence. His research also investigates how community composition shifts in response to environmental changes, using pairwise competition outcomes to predict species survival and community dynamics. Additionally, Gore studies the evolution of cooperation, showing how cooperative yeast populations can coexist with cheaters due to preferential access to resources, and extends these concepts to antibiotic resistance. Recently, he has explored the physics of learning in neural networks, aiming to understand the origin of neural scaling laws and the dynamics of high-dimensional training processes. His interdisciplinary approach combines experimental microbiology, theoretical modeling, and physics to address fundamental questions in ecology, evolution, and complex systems. Gore received his PhD from UC Berkeley in 2005, supported by a Hertz Fellowship, and has been recognized with numerous awards including the Schmidt Science Polymath Award, Allen Distinguished Investigator Award, NIH New Innovator Award, and Sloan Foundation Fellowship. He has also been active in community leadership, serving as co-Chair of the MIT Microbiology Program, US representative of the IUPAP Commission on Biological Physics, and an editor at PLOS Biology. Gore is passionate about teaching and mentoring, having received the Buechner Teaching Award and the UROP Faculty Mentor of the Year Award.

Research topics

• Computer Science
• Ecology
• Microbiology
• Biology
• Evolutionary biology

Selected publications

• Superposition unifies power-law training dynamics

  ArXiv.org · 2026-02-01

  articleOpen accessSenior author

  We investigate the role of feature superposition in the emergence of power-law training dynamics using a teacher-student framework. We first derive an analytic theory for training without superposition, establishing that the power-law training exponent depends on both the input data statistics and channel importance. Remarkably, we discover that a superposition bottleneck induces a transition to a universal power-law exponent of $\sim 1$, independent of data and channel statistics. This one over time training with superposition represents an up to tenfold acceleration compared to the purely sequential learning that takes place in the absence of superposition. Our finding that superposition leads to rapid training with a data-independent power law exponent may have important implications for a wide range of neural networks that employ superposition, including production-scale large language models.

  PublisherOA PDF

• Superposition unifies power-law training dynamics

  Open MIND · 2026-02-01

  preprintSenior author

  We investigate the role of feature superposition in the emergence of power-law training dynamics using a teacher-student framework. We first derive an analytic theory for training without superposition, establishing that the power-law training exponent depends on both the input data statistics and channel importance. Remarkably, we discover that a superposition bottleneck induces a transition to a universal power-law exponent of $\sim 1$, independent of data and channel statistics. This one over time training with superposition represents an up to tenfold acceleration compared to the purely sequential learning that takes place in the absence of superposition. Our finding that superposition leads to rapid training with a data-independent power law exponent may have important implications for a wide range of neural networks that employ superposition, including production-scale large language models.

  DOI

• Interspecies Interactions Drive Community-Level Selection in Microbial Coalescence

  bioRxiv (Cold Spring Harbor Laboratory) · 2026-02-14

  articleOpen accessSenior authorCorresponding

  Abstract It has long been debated in ecology whether communities behave as cohesive units or as loose collections of independent species. Here, we study this question in the context of community coalescence, the mixing of previously isolated communities, using bacterial microcosm experiments combined with ecological modeling. Our results demonstrate that interspecies interaction strength determines whether communities or species are the units of selection during coalescence. When interactions are moderate to strong, one parental community consistently outcompetes the other, indicating community-level selection. In contrast, under weak interactions, species fates are uncorrelated and the two communities contribute equally to the coalesced outcome, indicating the absence of community-level selection. These patterns extend to communities derived from natural samples with greater taxonomic diversity and richness. Furthermore, we identify two distinct regimes underlying community-level selection in experiments with different media conditions: an emergent regime in which collective dynamics shape outcomes that cannot be predicted from species traits alone, and a top-down regime where dominant species determine the winning community. Together, these results reconcile conflicting observations on community-level selection during community coalescence by demonstrating that communities behave as cohesive units only when interactions are sufficiently strong.

  PublisherDOI

• Repulsion from slow-diffusing nutrients improves microbial chemotaxis towards moving sources

  Nature Communications · 2026-04-04

  articleOpen accessSenior author

  Chemotaxis, or the following of chemical concentration gradients, is essential for microbes to locate nutrients. However, microbes often display paradoxical behaviors, such as Escherichia coli being repelled by several amino acids. Here, we explore chemotaxis towards a moving source and demonstrate that when multiple nutrients are released from the source repulsion from certain nutrients actually improves chemotaxis towards the source. Because a moving source leaves most of the nutrient plume behind it, simply following the concentration gradient results in aiming behind the source and potentially failing to intercept it. However, when attraction to a fast-diffusing nutrient and repulsion from a slow-diffusing nutrient are combined, motion in a new direction emerges and the chance of intercepting the source is increased up to six-fold. We demonstrate that this "differential strategy" is robust against numerous variations, including order-of-magnitude increases in the repellent release rate. Finally, we leverage existing data to show that E. coli is attracted to fast-diffusing amino acids and repelled by slow-diffusing ones, suggesting it may utilize a differential strategy and providing an explanation for its repulsion from these amino acids. Our results thus illuminate new possibilities in how microbes can integrate signals from multiple gradients to accomplish challenging chemotactic tasks.

  PublisherDOI

• Nutrient availability shapes the diversity and structure of microbial communities

  bioRxiv (Cold Spring Harbor Laboratory) · 2026-04-16

  articleSenior author

  Abstract Nutrients are key drivers of microbial community structure, yet we lack a data-driven quantitative framework linking nutrient environments to community assembly. Here, using controlled microcosm experiments, we systematically probe the effects of nutrient number, concentration, and type on community diversity and structure. To explain these effects, we construct a minimal consumer-resource model incorporating resource competition and cross-feeding. We find that cross-feeding network structure is critical: only shallow, wide networks — where several byproducts are produced from a few supplied nutrients in a few trophic layers — reproduce a linear increase in diversity with the number of supplied nutrients. We also perform new experiments varying nutrient concentration and reveal that diversity slowly decreases with increasing concentration. We explain this finding by incorporating consumption-dependent toxicity into the model, consistent with spent-media measurements. This extended model not only recapitulates virtually all observed patterns but makes an independent prediction: at high nutrient concentrations, communities should be enriched in bistable species pairs, which we confirm experimentally. Our work demonstrates that minimal data-driven consumer-resource frameworks — systematically constrained by experiments — can unify and predict a broad range of nutrient–community relationships.

  PublisherDOI

• Interspecies Interactions Drive Community-Level Selection in Microbial Coalescence

  Research Square · 2026-03-04

  preprintOpen access1st authorCorresponding
  PublisherOA PDFDOI

• Transition from global stability to multiple attractors in microcosms

  Research Square · 2025-11-20 · 1 citations

  preprintOpen access1st authorCorresponding
  PublisherOA PDFDOI

• Structured interactions explain the absence of keystone species in synthetic microcosms

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-04-05 · 2 citations

  preprintOpen accessSenior authorCorresponding

  Abstract In complex ecosystems, the loss of certain species can trigger a cascade of secondary extinctions and invasions. However, our understanding of the prevalence of these critical “keystone” species and the factors influencing their emergence remains limited. To address these questions, we experimentally assembled microcosms from 16 marine bacterial species and found that multiple extinctions and invasions were exceedingly rare upon removal of a species from the initial inoculation. This was true across eight different environments with either simple carbon sources (e.g., glucose) and more complex ones (e.g. glycogen). By employing a generalized Lotka-Volterra model, we could reproduce these results when interspecies interactions followed a hierarchical pattern, wherein species impacted strongly by one species were also more likely to experience strong impacts from others. Such a pattern naturally emerges due to observed variation in carrying capacities and growth rates. Furthermore, using both statistical inference and spent media experiments, we inferred interspecies interaction strengths and found them consistent with structured interactions. Our results suggest that the natural emergence of structured interactions may provide community resilience to extinctions.

  PublisherOA PDFDOI

• FOCUS: First Order Concentrated Updating Scheme

  ArXiv.org · 2025-01-21

  preprintOpen accessSenior author

  Large language models (LLMs) demonstrate remarkable performance, and improving their pre-training process appears to be key to enhancing their capabilities further. Based on the documented success of Adam, learning rate decay, and weight decay, we hypothesize that the pre-training loss landscape features a narrowing valley structure. Through experiments with synthetic loss functions, we discover that when gradient query noise is high relative to the valley's sharpness, Adam's performance falls behind that of Signum because Adam reduces the effective step size too drastically. This observation led us to develop FOCUS, an optimizer that enhances Signum by incorporating attraction toward moving averaged parameters, allowing it to handle noise better while maintaining larger step sizes. In training GPT-2, FOCUS proves to be more stable than Signum and faster than Adam. These results suggest that gradient noise may be an underappreciated limiting factor in LLM training, and FOCUS offers promising solutions.

  PublisherOA PDFDOI

• Superposition Yields Robust Neural Scaling

  ArXiv.org · 2025-05-15

  preprintOpen accessSenior author

  The success of today's large language models (LLMs) depends on the observation that larger models perform better. However, the origin of this neural scaling law, that loss decreases as a power law with model size, remains unclear. We propose that representation superposition, meaning that LLMs represent more features than they have dimensions, can be a key contributor to loss and cause neural scaling. Based on Anthropic's toy model, we use weight decay to control the degree of superposition, allowing us to systematically study how loss scales with model size. When superposition is weak, the loss follows a power law only if data feature frequencies are power-law distributed. In contrast, under strong superposition, the loss generically scales inversely with model dimension across a broad class of frequency distributions, due to geometric overlaps between representation vectors. We confirmed that open-sourced LLMs operate in the strong superposition regime and have loss scaling inversely with model dimension, and that the Chinchilla scaling laws are also consistent with this behavior. Our results identify representation superposition as a central driver of neural scaling laws, providing insights into questions like when neural scaling laws can be improved and when they will break down.

  PublisherOA PDFDOI

Recent grants

• Environmental modulation of microbial conflict and cooperation

  NIH · $2.3M · 2013–2022

• NIH Grant K99GM085279

  NIH · $90k · 2010

• CAREER: Exploration of Evolutionary Dynamics on Rugged Fitness Landscapes

  NSF · $600k · 2011–2016

• Early warning indicators of tipping points in biological systems

  NIH · $2.3M · 2012–2017

• NIH Grant R00GM085279

  NIH · $742k · 2014

Frequent coauthors

• Carlos Bustamante

  University of California, Berkeley

  49 shared

• Christoph Ratzke

  University of Tübingen

  48 shared

• Alfonso Pérez‐Escudero

  Université Toulouse III - Paul Sabatier

  44 shared

• Hyun-Seok Lee

  Massachusetts Institute of Technology

  44 shared

• Matthieu Barbier

  Centre de Coopération Internationale en Recherche Agronomique pour le Développement

  38 shared

• Martina Dal Bello

  Living Systems (United States)

  38 shared

• Zev Bryant

  Stanford University

  36 shared

• Daniel R. Amor

  Université Sorbonne Paris Nord

  35 shared

Labs

• Gore LaboratoryPI

Awards & honors

• Schmidt Science Polymath Award (2020)
• Allen Distinguished Investigator Award
• NIH New Innovator Award
• Sloan Foundation Fellowship
• NSF Career Award