Artificial Intelligence (AI) is transforming how scientific research is conducted, from hypothesis generation and data collection to analysis, writing, and peer review. In doing so, AI-driven acceleration is creating an emerging evaluation bottleneck in scientific knowledge production, where the generative capacity of science expands faster than the institutional capacity to evaluate, verify, and contest scientific claims, producing structural asymmetries in the mechanisms of scrutiny on which reliable and trustworthy knowledge depend. We develop a conceptual framework that explains how AI influences science through four analytical roles that interact with the institutional structures and evaluative processes of science. Specifically, AI can function as a telescope extending scientific observation, as a partner assisting scientific reasoning, as a paradigm reshaping evaluative standards, and as a disruptive force challenging established authority. AI can advance science when these roles strengthen testing and understanding, but it can undermine the scientific method when speed and scale outpace careful evaluation, causal explanation, and human judgment. We show that the evaluation bottleneck emerges from the interaction of these four roles across different stages of scientific practice. This paper examines how this bottleneck influences scientific practice across key stages of research and how it affects the reliability, explainability, and contestability of scientific knowledge. To address this evaluation bottleneck, we argue that scientific institutions must strengthen governance arrangements and evaluative mechanisms that sustain independent scrutiny, transparency, and replicability in AI-enabled science.

Counterfactual estimation over time is important in various applications, such as personalized medicine. However, time-dependent confounding bias in observational data still poses a significant challenge in achieving accurate and efficient estimation. We introduce causal autoencoding and treatment conditioning (CAETC), a novel method for this problem. Built on adversarial representation learning, our method leverages an autoencoding architecture to learn a partially invertible and treatment-invariant representation, where the outcome prediction task is cast as applying a treatment-specific conditioning on the representation. Our design is independent of the underlying sequence model and can be applied to existing architectures such as long short-term memories (LSTMs) or temporal convolution networks (TCNs). We conduct extensive experiments on synthetic, semi-synthetic, and real-world data to demonstrate that CAETC yields significant improvement in counterfactual estimation over existing methods.

Large language models (LLMs) require considerable computation and energy resources during training and deployment. While scaling laws for training have guided much recent progress, inference costs represent a significant and growing component of the overall resource burden, particularly for reasoning models. Existing compute-optimality characterizations that consider model size, dataset size and inference tokens in isolation or fixed combinations may overlook more efficient operating points. We introduce directed stochastic skill search (DS3), a general framework that represents inference as stochastic traversal over a learnt skill graph. From a simplified yet expressive instantiation, we derive closed-form expressions for task success and compute cost across a wide range of inference strategies-including chain-of-thought (CoT) and tree-of-thought (ToT)-enabling comparisons by task difficulty and model capability. We extend a prior graph framework of LLM training to include inference and bridge DS3 with empirical scaling laws. We theoretically recover observed patterns, including linear accuracy scaling with log-compute, variation in preferred inference strategies by task and capability, emergent behaviour elicited by reasoning despite parameter plateaus and both best-of-N and majority voting (MV) captured within one framework. By characterizing training-inference interdependencies, our framework deepens theoretical understanding and supports principled algorithmic design and resource allocation. This article is part of the discussion meeting issue 'Bits, neurons and qubits for sustainable AI'.

The growing integration of AI agents into human teams calls for a principled understanding of how collective intelligence emerges in hybrid systems. Recent frameworks clarify how attention, memory, and reasoning differences shape human-AI interaction at the individual and dyadic levels, but a formal account of how these differences scale to group-level dynamics is lacking. Most network science has examined either human-only or multi-agent AI-only systems, leaving open how its findings and parametrizations translate to hybrid groups. This chapter synthesizes network science, collective cognition, and multi-agent systems through the lens of attention, memory, and reasoning. We review how task environments, group topologies, agent-level processes, and incentive structures shape collective outcomes in human-only and AI-only networks, then examine how these results extend to hybrid settings, conceptualizing hybrid networks as heterogeneous human-AI nodes and links with distinct individual and transactive constraints. Our comparative analysis identifies which network effects are robust across agent types and which require revision, and highlights configurations that were peripheral in single-type traditions, such as human gatekeepers of AI sub-networks, but become structurally central in hybrid teams. Integrating a cognitive systems perspective with network science, we clarify how established exploration-exploitation and efficiency-redundancy trade-offs may operate differently in hybrid teams, and conclude with implications for organizational design, governance, and the responsible development of hybrid intelligence systems.

Controlled research on AI ideation typically compares independent agents, while field studies of human-AI collaboration sacrifice experimental control. We introduce a controlled, two-player extension of the Alternate Uses Test (AUT) that enables comparison of human-human and human-AI co-creation under matched interactive conditions, alongside calibrated non-interactive baselines. The platform supports decomposition of performance into three typically confounded factors: participant traits, partner perceptions, and content dynamics. An in-person pilot (N = 62) demonstrates its utility. Under matched time limits, originality with a GPT-4 partner is statistically equivalent to that with a human partner. Approach motivation (BAS Drive) moderates whether interactive partnership benefits originality, and self-reported cognitive outsourcing predicts lower originality specifically in human-human dyads. Prior exposure to highly creative ideas improves later performance, suggesting a "seeding" intervention. We release the platform, code, and dataset as a shared testbed for controlled studies of human-AI co-creation.

Large reasoning models (LRMs) have heterogeneous inference energy costs based on which model is used and how much it reasons. To reduce energy, it is important to choose the right LRM and operate it in the right way. As a result, the performance of systems that dispatch tasks to different individual LRMs depend on the balance between mean energy provisioning and stochastic fluctuations. The critical regime is the unique operating point at which neither auxiliary energy nor baseline energy is systematically wasted. Increasing baseline supply shifts the system toward persistent over-supply and baseline-energy waste, while reducing supply induces persistent reliance on auxiliary energy. Yet in this regime, performance remains volatility-limited and so a second-order characterization provides further insights that we develop. Here, performance is governed by how variability is absorbed across time, models, and execution choices. This perspective highlights variance-aware routing and dispatch as a principled design axis, and provides a theoretical basis for developing energy-aware model routing policies. We characterize routing behavior under a simple dispatch policy based on training- and inference-compute scaling laws for LRMs.

Hopfield networks and their generalizations have established deep connections among biological associative memories, statistical physics, and transformers. Yet most models treat retrieval as a fixed query-to-memory mapping, ignoring the role of external context in recall. In this work, we propose a two-stage associative memory architecture, wherein a context-gate subcircuit reshapes the retrieval energy landscape before and during recall. We show theoretically that context gating increases inter-memory separation while inducing sparsity, translating into exponential improvements in retrieval. Crucially, we prove that the system admits a unique self-consistent fixed point, revealing that the resulting retrieval state is driven by both a direct contextual bias and a second-order retrieval-gate feedback loop. We then bridge this theory to transformers; specifically, we evaluate a first-order approximation on Llama-3, confirming that in-context learning acts as context-gated retrieval. Native dynamics mirror our theory: context localizes a memory subspace, enabling the zero-shot query to cleanly discriminate. Ultimately, this framework provides a mechanistic link between associative memory theory and LLM phenomenology.

We introduce chaos-controlled Reservoir Computing (cc-RC) for living neural cultures: dynamically rich substrates of unique potential for adaptive computation. To account for intrinsic biological variability, cc-RC combines: (i) pre-training identification of each culture's dynamical signature and phase-portrait attractor; (ii) low-power optical chaos control to stabilize spontaneous and stimulus-evoked activity; (iii) readout training within this controlled regime. Across hundreds of neural samples, cc-RC enables robust learning and pattern classification, improving both accuracy and model longevity by approximately 300

More than a billion people communicate in Indic languages, yet the natural language processing infrastructure serving them remains fragmented and underdeveloped. The cause is structural: the field organizes its tools and benchmarks around individual languages or small subsets of genealogical language families, building separate analyzers, parsers, and datasets for each language and starting over for the next. This overlooks a deep regularity. Through more than two millennia of convergence around Sanskrit, Indic languages came to share a morphosyntactic architecture formalized in Pānini's grammar, the Astādhyāyī. This cuts across genealogical lines, uniting languages through a common framework. We argue that this Pāninian framework supplies a unifying computational architecture the field has lacked, and that benchmarks grounded explicitly in it would make Indic language systems more accurate, more data-efficient, and more transferable, effectively merging many apparently disparate and sparse Indic language resources into a single high-resource metalanguage bedrock. We propose a four-part benchmark suite to render this shared architecture explicit, measurable, and ready to be leveraged for practical applications. Moreover, we underscore the question it raises for interpretability research: whether neural models trained on these languages come to represent Pānini's categories on their own.

Divergent thinking is a crucial aspect of creativity, yet large language models (LLMs) tend to consistently generate similar responses to open-ended questions, in what has been termed the artificial hivemind effect. Here, we introduce CreativityNeuro, a data-free method for enhancing divergent thinking in LLMs via contrastive weight steering. We evaluate our method across multiple creativity assessments and report several main findings. On the Divergent Association Task (DAT), a vocabulary-space creativity test, CreativityNeuro improves performance by up to 14 human percentile points. Next, in a large-scale human evaluation (N=720) on the Alternative Uses Test (AUT) and the Task Task, CreativityNeuro achieves significant improvements in originality, surprise, and creativity, transferring to longer-form and more open-ended tasks. Importantly, we find that across all three tasks, CreativityNeuro demonstrably reduces measures of mode collapse. Moreover, activation steering achieves comparable performance to CreativityNeuro on the DAT, but it does not transfer to the AUT and Task Task, demonstrating the effectiveness of weight-space steering in generalizing to unseen tasks. In conclusion, CreativityNeuro improves divergent thinking and reduces mode collapse without requiring behavioral data, re-training, or gradient-based fine-tuning, providing a straightforward way to enhance LLM performance in creative domains.

Fault-tolerant quantum computation demands extremely low logical error rates, yet superconducting qubit arrays are subject to radiation-induced correlated noise arising from cosmic-ray muon-generated quasiparticles. The quasiparticle density is unknown and time-varying, resulting in a mismatch between the true noise statistics and the priors assumed by standard decoders, and consequently, degraded logical performance. We formalize joint noise sensing and decoding using syndrome measurements by modeling the QP density as a latent variable, which governs correlation in physical errors and syndrome measurements. Starting from a variational expectation–maximization approach, we derive an iterative algorithm that alternates between QP density estimation and syndrome-based decoding under the updated noise model. Simulations of surface-code and bivariate bicycle quantum memory under radiation-induced correlated noise demonstrate a measurable reduction in logical error probability relative to baseline decoding with a uniform prior. Beyond improved decoding performance, the inferred QP density provides diagnostic information relevant to device characterization, shielding, and chip design. These results indicate that integrating physical noise estimation into decoding can mitigate correlated noise effects and relax effective error-rate requirements for fault-tolerant quantum computation.

We consider federated learning of linearly-parameterized nonlinear systems. We establish theoretical guarantees on the effectiveness of federated nonlinear system identification compared to centralized approaches, demonstrating that the convergence rate improves as the number of clients increases. Although the convergence rates in the linear and nonlinear cases differ only by a constant, this constant depends on the feature map ϕ, which can be carefully chosen in the nonlinear setting to increase excitation and improve performance. We experimentally validate our theory in physical settings where client devices are driven by i.i.d. control inputs and control policies exhibiting i.i.d. random perturbations, ensuring non-active exploration. Experiments use trajectories from nonlinear dynamical systems characterized by real-analytic feature functions, including polynomial and trigonometric components, representative of physical systems including pendulum and quadrotor dynamics. We analyze the convergence behavior of the proposed method under varying noise levels and data distributions. Results show that federated learning consistently improves convergence of any individual client as the number of participating clients increases.

The hyperscaling of data and parameter count in transformer models is yielding diminishing performance improvement, especially when weighed against training costs. Such plateauing underlines a growing need for more efficient finetuning and inference, without sacrificing performance. This is particularly pressing for multimodal learning, where the overhead of processing multimodal tokens alongside language data often limits the practical viability of these systems. In parallel, advances in representation learning and interpretability have deepened our understanding of how such models process and encode information. Notably, recent work has uncovered implicit cross-modal alignment in the deeper layers of large pretrained models. Interestingly, this aligns with our own observations that models naturally defer most cross-modal token interactions to deeper stages of computation. Building on this, we propose a simple modification. Instead of concatenation with the language prompt at the start, we insert multimodal tokens directly into the middle, allowing them to entirely bypass the early layers. Our results with diverse modalities: 1) LLaVA & BLIP for vision, 2) LTU for audio, and 3) MoLCA for molecular data, indicate that our method reduces computational costs during both training and inference, while at the very least, preserving, if not surpassing the performance of existing baselines. Our work has important implications for scaling and composing pretrained models in a resource-efficient manner.

Information lattice learning (ILL) learns interpretable rules of a signal by alternately projecting the signal onto a partition lattice that encodes a hierarchy of abstractions and lifting selected rules back to the signal domain. When the signal is a probability mass function, we show the probabilistic rules learned by ILL admit a natural probabilistic graphical model (PGM) interpretation and develop this interpretation in detail. A partition in ILL induces a deterministic quotient variable, and a rule is the marginal law of that quotient variable. A rule set is therefore a collection of marginal constraints over interpretable abstractions. General lifting is the feasible family of all joint distributions satisfying those constraints, while special lifting chooses a maximum-ignorance reconstruction, implemented in ILL by an L2 uniformity principle closely related to maximum entropy. Under a Shannon-entropy lifting, the same constraints yield a log-linear factor graph whose factors are indexed by learned abstractions. The information lattice itself, however, is not a Bayesian network: its edges encode refinement and coarsening of abstractions, not conditional dependence. Thus ILL is best viewed as structure learning for interpretable constraint-based factor graphs over quotient variables. This view clarifies how ILL relates to graphical models and maximum entropy models, while suggesting new directions for inference, identifiability, and hybrid symbolic-probabilistic learning.

Agentic frameworks are the software layer through which AI agents act in the world. Existing safety methods intervene on the model and therefore remain conditional on unverifiable properties of learned behavior. We introduce containment verification, which locates safety guarantees in the agentic framework itself. Under havoc oracle semantics, the AI is modeled as an unconstrained oracle ranging over the entire typed action space, and the verified containment layer must enforce the boundary policy for every possible AI output. For boundary-enforceable properties, expressed over modeled boundary events, action arguments, and state, we prove a universal guarantee by forward-simulation refinement and mechanize it in Dafny. We instantiate the paradigm by verifying PocketFlow, a minimalist agentic LLM framework, and use an agentic synthesis pipeline to generate the specification, operational model, and refinement proof under an information barrier against tautological specifications. To our knowledge, this is the first deductive formal verification of an agentic framework, and its guarantee is invariant to model capability over the modeled typed action boundary.

Low-Rank Adaptation (LoRA) has become ubiquitous for efficiently fine-tuning foundation models. However, federated fine-tuning using LoRA is challenging due to suboptimal updates arising from traditional federated averaging of individual adapters. Existing solutions either incur prohibitively high communication cost that scales linearly with the number of clients or suffer from performance degradation due to limited expressivity. We introduce Federated Silver Bullet (Fed-SB), a novel approach for federated fine-tuning of LLMs using LoRA-SB, a recently proposed low-rank adaptation method. LoRA-SB optimally aligns the optimization trajectory with the ideal low-rank full fine-tuning projection by learning a small square matrix (R) between adapters B and A, keeping other components fixed. Direct averaging of R guarantees exact updates, substantially reducing communication cost, which remains independent of the number of clients, and enables scalability. Fed-SB achieves state-of-the-art performance across commonsense reasoning, arithmetic reasoning, and language inference tasks while reducing communication costs by up to 230x. In private settings, Fed-SB further improves performance by (1) reducing trainable parameters, thereby lowering the noise required for differential privacy and (2) avoiding noise amplification introduced by other methods. Overall, Fed-SB establishes a new Pareto frontier in the tradeoff between communication and performance, offering an efficient and scalable solution for both private and non-private federated fine-tuning. Our code is publicly available at https://github.com/CERT-Lab/fed-sb.

Rigorous economic models can take months to construct, yet energy crises demand decisions from policymakers within days or even hours. Any disruption in energy markets is not isolated but rapidly disseminates through interlinked global systems. Off-the-shelf models that already exist typically focus only on limited aspects of the system and are distributed across research groups, programming languages, software architectures not designed for model integration, and incompatible formats. Integrating these models manually can take longer than the crisis itself, forcing analysts to rely on whichever models are easiest to connect and leaving consequential scenarios unexplored. Policymakers must make rapid decisions with obstructed and limited information. We show that large language models can perform the critical integration directly. The system constructs internally consistent scenarios, translates assumptions into model-specific inputs, executes existing economic and physical models in dependency order, and synthesizes outputs tailored to policymakers. The language model generates no quantitative results: every reported value is reproduced directly from an underlying model run, remains traceable to its source and is subject to analyst approval at each stage. We develop a LLM framework that coordinates 16 models of oil, natural gas, shipping, water, helium, fertilizer and macroeconomic equilibrium. The framework is applied across five scenarios to assess the 2026 closure of the Strait of Hormuz and refreshed weekly for eight weeks as events on the ground continued to unfold. By linking models that already exist and reading them as a suite rather than in isolation, this architecture mobilizes distributed scientific models rapidly during energy and geopolitical disruptions while keeping any single model's assumptions from driving the conclusion.

Large language models (LLMs) hallucinate: they produce fluent outputs that are factually incorrect. We present a geometric dynamical systems framework in which hallucinations arise from task-dependent basin structure in latent space. Using autoregressive hidden-state trajectories across multiple open-source models and benchmarks, we find that separability is strongly task-dependent rather than universal: factoid settings can show clearer basin separation, whereas summarization and misconception-heavy settings are typically less stable and often overlap. We formalize this behavior with task-complexity and multi-basin theorems, characterize basin emergence in L-layer transformers, and show that geometry-aware steering can reduce hallucination probability without retraining.

We study distributed experimental design with multiple candidate experiments, where local sites possess only partial information and transmit design recommendations to a fusion center. Unlike centralized design, in which the experiment that maximizes expected information gain can be selected directly, distributed design requires combining heterogeneous and potentially conflicting local recommendations. Formulating as a multi-class Bayes fusion problem, centralized oracle design is treated as an unknown label and each site is characterized by a local recommendation mechanism. The proposed fusion rule minimizes posterior expected information regret, rather than merely maximizing the number of local votes or the posterior probability (MAP) of the oracle label. This distinction is essential because different incorrect experimental choices may incur different losses in information gain. We show that majority vote is optimal only under restrictive symmetry assumptions and can otherwise be strictly suboptimal. Regret-weighted multi-class Chernoff bounds are derived to identify the pairwise separations governing distributed design performance. Numerical studies identify two operational regimes: MAP is effective when oracle-label accuracy and information regret are aligned, while regret-weighted Bayes fusion reduces information loss when the most probable oracle label is not the lowest-regret decision.

Distillation via sampling reasoning traces exposes closed-source frontier models to adversarial third parties who can bypass their guardrails and misappropriate their capabilities. Antidistillation methods aim to address this by poisoning reasoning traces to hinder student model learning while preserving teacher performance. However, current methods overlook detectability, both semantic and syntactic, which erodes trust in the teacher's outputs and signals the defense's presence to adversaries. We address this gap by formulating antidistillation as a Stackelberg game whose constraint set explicitly encodes detectability, and show that perturbing sparingly offers an effective, less detectable alternative to poisoning the full trace. Drawing on mechanistic interpretability, we identify thought anchors, sentences with disproportionate counterfactual influence on model outputs, as a principled sparse target: critical to reasoning yet minimally detectable. We instantiate this in TraceGuard, a training-free, black-box proof-of-concept that locates thought anchors via branching-token detection and poisons them to degrade student distillation while preserving trace coherence.